{"source":"AIMap","schema":"learn-algorithms/v1","updated":"2026-10-22","note":{"zh":"year 取原始论文首次公开年份（arXiv 首版或正式发表年）；paper.check=verified 表示标题、年份、链接已通过 arXiv API / Crossref / 官方页面核实。","en":"year = first public version of the original paper. paper.check=verified means title/year/link were checked via arXiv API, Crossref or the official page."},"categories":[{"id":"classic-ml","name":{"zh":"经典机器学习","en":"Classical Machine Learning"},"description":{"zh":"深度学习之前就成熟的统计学习方法，表格数据和小数据场景至今仍是主力。","en":"Statistical learning methods that predate deep learning and still dominate tabular and small-data problems."},"count":18,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/classic-ml/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/classic-ml/"}},{"id":"dl-foundations","name":{"zh":"深度学习基础","en":"Deep Learning Foundations"},"description":{"zh":"神经网络如何被训练：反向传播、优化器、归一化、激活函数与经典网络结构。","en":"How neural nets are trained: backprop, optimizers, normalization, activations and classic architectures."},"count":23,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/dl-foundations/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/dl-foundations/"}},{"id":"transformer","name":{"zh":"Transformer 与注意力","en":"Transformers & Attention"},"description":{"zh":"大模型的骨架：注意力、位置编码、分词，以及 FlashAttention、GQA/MLA、长上下文等变体。","en":"The backbone of LLMs: attention, positional encodings, tokenization and variants like FlashAttention, GQA/MLA and long context."},"count":18,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/transformer/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/transformer/"}},{"id":"moe","name":{"zh":"混合专家 MoE","en":"Mixture of Experts"},"description":{"zh":"只激活部分参数的稀疏模型，让总参数量大幅增加而计算量不按比例增长。","en":"Sparse models that activate only some parameters, growing capacity without proportional compute."},"count":6,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/moe/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/moe/"}},{"id":"new-arch","name":{"zh":"新架构：SSM / 线性注意力","en":"New Architectures: SSM & Linear Attention"},"description":{"zh":"Mamba、RWKV、线性注意力、混合架构与扩散语言模型等 Transformer 之外的路线。","en":"Post-Transformer directions: Mamba, RWKV, linear attention, hybrids and diffusion language models."},"count":11,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/new-arch/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/new-arch/"}},{"id":"pretraining","name":{"zh":"预训练、缩放与分布式训练","en":"Pretraining, Scaling & Distributed Training"},"description":{"zh":"缩放定律、超参迁移、混合精度与并行训练，决定一个大模型怎么被“炼”出来。","en":"Scaling laws, hyperparameter transfer, mixed precision and parallelism — how frontier models get trained."},"count":9,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/pretraining/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/pretraining/"}},{"id":"alignment","name":{"zh":"对齐与偏好优化","en":"Alignment & Preference Optimization"},"description":{"zh":"让模型听话、有用、无害：RLHF、DPO、宪法式 AI、GRPO 与可验证奖励强化学习。","en":"Making models helpful and harmless: RLHF, DPO, Constitutional AI, GRPO and RL with verifiable rewards."},"count":10,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/alignment/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/alignment/"}},{"id":"reasoning","name":{"zh":"推理模型与测试时计算","en":"Reasoning & Test-Time Compute"},"description":{"zh":"思维链、自洽性、过程奖励与“多想一会儿”的测试时扩展，是 o1 / R1 类推理模型的基础。","en":"Chain of thought, self-consistency, process rewards and test-time scaling behind o1/R1-style reasoners."},"count":9,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/reasoning/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/reasoning/"}},{"id":"finetuning","name":{"zh":"微调与参数高效训练","en":"Fine-tuning & PEFT"},"description":{"zh":"用少量参数或数据把通用模型改造成专用模型：LoRA、QLoRA、指令微调、模型合并。","en":"Adapting general models cheaply: LoRA, QLoRA, instruction tuning and model merging."},"count":8,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/finetuning/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/finetuning/"}},{"id":"compression","name":{"zh":"量化、剪枝与蒸馏","en":"Quantization, Pruning & Distillation"},"description":{"zh":"把大模型变小、变快、变省显存，让它能跑在单卡、笔记本甚至手机上。","en":"Making models smaller, faster and lighter so they run on one GPU, a laptop or a phone."},"count":7,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/compression/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/compression/"}},{"id":"inference","name":{"zh":"推理加速与服务","en":"Inference Acceleration & Serving"},"description":{"zh":"KV Cache、推测解码、PagedAttention、连续批处理等让大模型服务又快又便宜的技术。","en":"KV cache, speculative decoding, PagedAttention, continuous batching — making LLM serving fast and cheap."},"count":10,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/inference/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/inference/"}},{"id":"generative","name":{"zh":"扩散、流匹配与生成模型","en":"Diffusion, Flow Matching & Generative Models"},"description":{"zh":"VAE、GAN、扩散模型、流匹配与 DiT：图像、视频、音频生成背后的算法。","en":"VAEs, GANs, diffusion, flow matching and DiT — the algorithms behind image, video and audio generation."},"count":15,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/generative/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/generative/"}},{"id":"multimodal","name":{"zh":"多模态与视觉","en":"Multimodal & Vision"},"description":{"zh":"ViT、CLIP、视觉语言模型、语音识别、分割与 3D 重建，让 AI 看、听、理解世界。","en":"ViT, CLIP, VLMs, speech, segmentation and 3D reconstruction — letting AI see and hear."},"count":15,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/multimodal/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/multimodal/"}},{"id":"retrieval","name":{"zh":"RAG 与向量检索","en":"RAG & Vector Retrieval"},"description":{"zh":"嵌入、向量索引、稠密检索、重排与 GraphRAG：给大模型接上外部知识。","en":"Embeddings, vector indexes, dense retrieval, reranking and GraphRAG — connecting LLMs to knowledge."},"count":13,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/retrieval/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/retrieval/"}},{"id":"agents","name":{"zh":"Agent 与工具调用","en":"Agents & Tool Use"},"description":{"zh":"ReAct、工具调用、MCP、反思与记忆：让模型能规划并操作外部世界。","en":"ReAct, tool calling, MCP, reflection and memory — letting models plan and act."},"count":9,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/agents/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/agents/"}},{"id":"rl","name":{"zh":"强化学习","en":"Reinforcement Learning"},"description":{"zh":"从 Q-learning 到 PPO、AlphaZero、MuZero：通过试错和奖励学习决策。","en":"From Q-learning to PPO, AlphaZero and MuZero — learning decisions from trial and reward."},"count":12,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/rl/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/rl/"}},{"id":"world-models","name":{"zh":"世界模型与自监督表征","en":"World Models & Self-Supervised Learning"},"description":{"zh":"在脑中模拟世界并预测未来：World Models、Dreamer、JEPA、Genie，以及对比学习、MAE 等自监督方法。","en":"Simulating and predicting the world: World Models, Dreamer, JEPA, Genie, plus self-supervised learning like contrastive learning and MAE."},"count":7,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/world-models/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/world-models/"}},{"id":"graph-science","name":{"zh":"图神经网络与科学 AI","en":"Graph Neural Networks & AI for Science"},"description":{"zh":"GCN、GAT 等图模型，以及 AlphaFold、AlphaGeometry 等改变科学研究的系统。","en":"Graph models like GCN and GAT, plus science systems like AlphaFold and AlphaGeometry."},"count":3,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/graph-science/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/graph-science/"}},{"id":"interpretability","name":{"zh":"可解释性","en":"Interpretability"},"description":{"zh":"打开黑箱：归因、探针、叠加、稀疏自编码器、电路追踪与激活引导。","en":"Opening the black box: attribution, probing, superposition, sparse autoencoders, circuit tracing and steering."},"count":9,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/interpretability/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/interpretability/"}},{"id":"safety","name":{"zh":"安全、评测与治理技术","en":"Safety, Evaluation & Security"},"description":{"zh":"对抗样本、越狱、提示注入、红队、水印、差分隐私与可扩展监督。","en":"Adversarial examples, jailbreaks, prompt injection, red teaming, watermarking, differential privacy and scalable oversight."},"count":11,"pages":{"en":"https://aimap.page/en/learn/algorithms/c/safety/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/c/safety/"}}],"algorithms":[{"id":"linear-regression","name":{"zh":"线性回归","en":"Linear Regression"},"aka":[],"category":"classic-ml","tier":"core","year":null,"difficulty":1,"summary":{"zh":"用一条直线（超平面）拟合输入与输出的关系。","en":"Fits a line (hyperplane) between inputs and a numeric output."},"explain":{"zh":"给每个特征一个权重，加起来就是预测值；训练就是找一组让预测误差平方和最小的权重。它是理解“损失函数”和“梯度下降”的最好起点。","en":"Each feature gets a weight and the weighted sum is the prediction; training finds weights minimizing squared error. It is the best entry point to loss functions and gradient descent."},"paper":null,"verify":{"status":"partial","checkedOn":"2026-10-21"},"prereqs":[],"usedBy":["logistic-regression","pca","sgd","lasso"],"related":["logistic-regression","sgd"],"usedIn":{"zh":"销量预测、房价估计、A/B 实验分析","en":"Sales forecasting, price estimation, A/B analysis"},"resources":[{"title":"scikit-learn 用户指南：线性模型","url":"https://scikit-learn.org/stable/modules/linear_model.html","type":"docs","lang":"en"},{"title":"An Introduction to Statistical Learning（ISLP，免费）","url":"https://www.statlearning.com/","type":"book","lang":"en"},{"title":"Stanford CS229 机器学习","url":"https://cs229.stanford.edu/","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/linear-regression/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/linear-regression/"}},{"id":"logistic-regression","name":{"zh":"逻辑回归","en":"Logistic Regression"},"aka":[],"category":"classic-ml","tier":"core","year":1958,"difficulty":1,"summary":{"zh":"用 S 形函数把线性得分变成概率，做二分类。","en":"Turns a linear score into a probability with a sigmoid for binary classification."},"explain":{"zh":"先像线性回归一样算一个得分，再用 sigmoid 压到 0~1 之间当作概率。用交叉熵损失训练，神经网络最后一层的 softmax 分类其实就是它的多类版本。","en":"Compute a linear score, squash it into 0–1 with a sigmoid, and train with cross-entropy. A neural network's final softmax layer is its multi-class generalization."},"paper":{"title":"The Regression Analysis of Binary Sequences","authors":["Cox et al."],"venue":"Journal of the Royal Statistical Society Series B: Statistical Methodology","year":1958,"doi":"10.1111/j.2517-6161.1958.tb00292.x","url":"https://doi.org/10.1111/j.2517-6161.1958.tb00292.x","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["linear-regression"],"usedBy":["svm","reward-model","knowledge-distillation","contrastive-learning","membership-inference"],"related":["perceptron","svm"],"usedIn":{"zh":"信用评分、点击率预估、医学风险模型","en":"Credit scoring, CTR prediction, clinical risk models"},"resources":[{"title":"scikit-learn 用户指南：线性模型","url":"https://scikit-learn.org/stable/modules/linear_model.html","type":"docs","lang":"en"},{"title":"An Introduction to Statistical Learning（ISLP，免费）","url":"https://www.statlearning.com/","type":"book","lang":"en"},{"title":"Stanford CS229 机器学习","url":"https://cs229.stanford.edu/","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/logistic-regression/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/logistic-regression/"}},{"id":"knn","name":{"zh":"K 近邻","en":"k-Nearest Neighbors"},"aka":[],"category":"classic-ml","tier":"core","year":1967,"difficulty":1,"summary":{"zh":"看离你最近的 k 个样本是什么，你就是什么。","en":"Predicts by majority vote of the k closest training examples."},"explain":{"zh":"不训练模型，只把数据存下来；预测时找出距离最近的 k 个邻居投票。它也是今天向量检索“找最相似”思想的祖先。","en":"There is no training: store the data and at prediction time let the k nearest neighbors vote. It is the ancestor of today's vector similarity search."},"paper":{"title":"Nearest neighbor pattern classification","authors":["Cover et al."],"venue":"IEEE Transactions on Information Theory","year":1967,"doi":"10.1109/TIT.1967.1053964","url":"https://doi.org/10.1109/TIT.1967.1053964","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":[],"usedBy":["hnsw","product-quantization"],"related":["hnsw","kmeans"],"usedIn":{"zh":"推荐、异常检测、向量检索的基线","en":"Recommendation, anomaly detection, vector-search baselines"},"resources":[{"title":"scikit-learn 用户指南：最近邻","url":"https://scikit-learn.org/stable/modules/neighbors.html","type":"docs","lang":"en"},{"title":"An Introduction to Statistical Learning（ISLP，免费）","url":"https://www.statlearning.com/","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/knn/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/knn/"}},{"id":"decision-tree","name":{"zh":"决策树（CART）","en":"Decision Trees (CART)"},"aka":[],"category":"classic-ml","tier":"core","year":null,"difficulty":1,"summary":{"zh":"用一连串“是/否”问题把数据切分成纯净的小组。","en":"Splits data with a sequence of yes/no questions into pure groups."},"explain":{"zh":"每一步选一个最能区分标签的特征和阈值把数据一分为二，直到叶子足够纯。单棵树易懂但容易过拟合，因此常被集成成随机森林或梯度提升树。","en":"At each step pick the feature/threshold that best separates labels and split, until leaves are pure. Single trees are interpretable but overfit, so they are usually ensembled."},"paper":null,"verify":{"status":"partial","checkedOn":"2026-10-21"},"prereqs":[],"usedBy":["random-forest","gbdt","adaboost"],"related":["random-forest","gbdt"],"usedIn":{"zh":"规则提取、风控、可解释模型","en":"Rule extraction, risk control, interpretable models"},"resources":[{"title":"scikit-learn 用户指南：决策树","url":"https://scikit-learn.org/stable/modules/tree.html","type":"docs","lang":"en"},{"title":"An Introduction to Statistical Learning（ISLP，免费）","url":"https://www.statlearning.com/","type":"book","lang":"en"},{"title":"StatQuest 视频索引","url":"https://statquest.org/video-index/","type":"video","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/decision-tree/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/decision-tree/"}},{"id":"random-forest","name":{"zh":"随机森林","en":"Random Forest"},"aka":[],"category":"classic-ml","tier":"core","year":2001,"difficulty":1,"summary":{"zh":"训练很多棵随机化的决策树，让它们投票。","en":"Trains many randomized decision trees and lets them vote."},"explain":{"zh":"每棵树只看随机抽取的样本和特征，彼此差异大；最后取平均或投票，误差相互抵消。几乎不用调参就很稳，是表格数据的强基线。","en":"Each tree sees a random sample of rows and features, so trees differ; averaging cancels their errors. It is robust with little tuning and a strong tabular baseline."},"paper":{"title":"Random Forests","authors":["Breiman et al."],"venue":"Machine Learning","year":2001,"doi":"10.1023/A:1010933404324","url":"https://doi.org/10.1023/A:1010933404324","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["decision-tree"],"usedBy":[],"related":["gbdt","xgboost"],"usedIn":{"zh":"表格分类回归、特征重要性分析","en":"Tabular classification/regression, feature importance"},"resources":[{"title":"scikit-learn 用户指南：集成方法","url":"https://scikit-learn.org/stable/modules/ensemble.html","type":"docs","lang":"en"},{"title":"An Introduction to Statistical Learning（ISLP，免费）","url":"https://www.statlearning.com/","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/random-forest/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/random-forest/"}},{"id":"gbdt","name":{"zh":"梯度提升树","en":"Gradient Boosting (GBDT)"},"aka":[],"category":"classic-ml","tier":"core","year":2001,"difficulty":2,"summary":{"zh":"一棵接一棵地训练小树，每棵专门修正前面的错误。","en":"Adds small trees one by one, each fixing the previous ensemble's errors."},"explain":{"zh":"把“沿梯度下降”的思想用在函数空间：每一轮拟合当前模型残差（负梯度）的一棵小树，再加到模型里。它是 XGBoost、LightGBM 的理论基础。","en":"Gradient descent in function space: each round fits a small tree to the residuals (negative gradient) and adds it in. It underlies XGBoost and LightGBM."},"paper":{"title":"Greedy function approximation: A gradient boosting machine.","authors":["Friedman et al."],"venue":"The Annals of Statistics","year":2001,"doi":"10.1214/aos/1013203451","url":"https://doi.org/10.1214/aos/1013203451","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["decision-tree","sgd"],"usedBy":["xgboost","lightgbm"],"related":["xgboost","lightgbm","random-forest","adaboost"],"usedIn":{"zh":"搜索排序、风控、Kaggle 表格赛","en":"Search ranking, risk scoring, Kaggle tabular competitions"},"resources":[{"title":"scikit-learn 用户指南：集成方法","url":"https://scikit-learn.org/stable/modules/ensemble.html","type":"docs","lang":"en"},{"title":"The Elements of Statistical Learning（免费）","url":"https://hastie.su.domains/ElemStatLearn/","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/gbdt/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/gbdt/"}},{"id":"xgboost","name":{"zh":"XGBoost","en":"XGBoost"},"aka":[],"category":"classic-ml","tier":"core","year":2016,"difficulty":2,"summary":{"zh":"工程化到极致的梯度提升树系统。","en":"A highly engineered, scalable gradient-boosted tree system."},"explain":{"zh":"在 GBDT 上加入正则化、二阶梯度、稀疏感知和并行/分布式计算，又快又准。多年来是表格数据竞赛的冠军常客。","en":"Adds regularization, second-order gradients, sparsity-awareness and parallel/distributed training on top of GBDT. It long dominated tabular competitions."},"paper":{"title":"XGBoost: A Scalable Tree Boosting System","authors":["Tianqi Chen","Carlos Guestrin"],"firstPublished":"2016-03-09","year":2016,"arxiv":"1603.02754","url":"https://arxiv.org/abs/1603.02754","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["gbdt"],"usedBy":[],"related":["lightgbm"],"usedIn":{"zh":"表格预测、广告点击率、风控","en":"Tabular prediction, ad CTR, risk"},"resources":[{"title":"XGBoost 官方文档","url":"https://xgboost.readthedocs.io/en/stable/","type":"docs","lang":"en"},{"title":"scikit-learn 用户指南：集成方法","url":"https://scikit-learn.org/stable/modules/ensemble.html","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/xgboost/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/xgboost/"}},{"id":"lightgbm","name":{"zh":"LightGBM","en":"LightGBM"},"aka":[],"category":"classic-ml","tier":"extended","year":2017,"difficulty":2,"summary":{"zh":"用直方图和按叶生长加速的梯度提升树。","en":"Gradient boosting accelerated by histograms and leaf-wise growth."},"explain":{"zh":"把连续特征分桶成直方图、只对梯度大的样本重点采样（GOSS）、合并互斥特征（EFB），训练速度和内存都大幅改善。","en":"Bins features into histograms, samples by gradient size (GOSS) and bundles exclusive features (EFB), cutting time and memory."},"paper":{"title":"LightGBM: A Highly Efficient Gradient Boosting Decision Tree","authors":["Ke et al."],"venue":"NeurIPS 2017","year":2017,"url":"https://papers.nips.cc/paper/2017/hash/6449f44a102fde848669bdd9eb6b76fa-Abstract.html","source":"官方页面 HTTP/WebFetch 核实","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["gbdt"],"usedBy":[],"related":["xgboost"],"usedIn":{"zh":"大规模表格学习、推荐排序","en":"Large-scale tabular learning, ranking"},"resources":[{"title":"LightGBM 官方文档","url":"https://lightgbm.readthedocs.io/en/stable/","type":"docs","lang":"en"},{"title":"microsoft/LightGBM（官方仓库）","url":"https://github.com/microsoft/LightGBM","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/lightgbm/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/lightgbm/"}},{"id":"svm","name":{"zh":"支持向量机","en":"Support Vector Machine"},"aka":[],"category":"classic-ml","tier":"core","year":1995,"difficulty":2,"summary":{"zh":"找一条离两类样本都尽量远的分界线。","en":"Finds the decision boundary with the widest margin between classes."},"explain":{"zh":"只由少数“支持向量”决定边界；配合核技巧可以在高维空间里画出弯曲的分界。深度学习普及前，它是图像和文本分类的主力。","en":"The boundary depends only on a few support vectors; the kernel trick yields nonlinear boundaries. Before deep learning it was the workhorse for vision and text classification."},"paper":{"title":"Support-vector networks","authors":["Cortes et al."],"venue":"Machine Learning","year":1995,"doi":"10.1007/BF00994018","url":"https://doi.org/10.1007/BF00994018","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["logistic-regression"],"usedBy":[],"related":["perceptron"],"usedIn":{"zh":"文本分类、小样本分类","en":"Text classification, small-data classification"},"resources":[{"title":"scikit-learn 用户指南：支持向量机","url":"https://scikit-learn.org/stable/modules/svm.html","type":"docs","lang":"en"},{"title":"Stanford CS229 机器学习","url":"https://cs229.stanford.edu/","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/svm/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/svm/"}},{"id":"kmeans","name":{"zh":"K 均值聚类","en":"k-Means Clustering"},"aka":[],"category":"classic-ml","tier":"core","year":1982,"difficulty":1,"summary":{"zh":"反复“分组—求中心”，把数据分成 k 簇。","en":"Alternates assigning points and recomputing centers to form k clusters."},"explain":{"zh":"随机放 k 个中心，每个点归到最近的中心，再把中心移到组内平均位置，直到不再变化。向量数据库里的 IVF 索引和乘积量化都在用它。","en":"Place k centers, assign points to the nearest, move centers to the mean, repeat. IVF indexes and product quantization in vector databases rely on it."},"paper":{"title":"Least squares quantization in PCM","authors":["Lloyd et al."],"venue":"IEEE Transactions on Information Theory","year":1982,"doi":"10.1109/TIT.1982.1056489","url":"https://doi.org/10.1109/TIT.1982.1056489","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":[],"usedBy":["em","product-quantization","dbscan"],"related":["em","product-quantization","pca"],"usedIn":{"zh":"用户分群、向量索引（IVF）、图像压缩","en":"Customer segmentation, IVF indexes, color quantization"},"resources":[{"title":"scikit-learn 用户指南：聚类","url":"https://scikit-learn.org/stable/modules/clustering.html","type":"docs","lang":"en"},{"title":"An Introduction to Statistical Learning（ISLP，免费）","url":"https://www.statlearning.com/","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/kmeans/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/kmeans/"}},{"id":"pca","name":{"zh":"主成分分析","en":"Principal Component Analysis (PCA)"},"aka":[],"category":"classic-ml","tier":"core","year":1901,"difficulty":1,"summary":{"zh":"找出数据变化最大的几个方向来降维。","en":"Reduces dimensions by keeping directions of greatest variance."},"explain":{"zh":"把数据旋转到一组新坐标轴上，第一轴方差最大，依次递减；只保留前几轴就能压缩数据、去噪、可视化。LoRA 的“低秩”直觉也与此相通。","en":"Rotates data onto axes ordered by variance; keeping the top few compresses, denoises and visualizes. The low-rank intuition behind LoRA is related."},"paper":{"title":"LIII. On lines and planes of closest fit to systems of points in space","authors":["Pearson et al."],"venue":"The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science","year":1901,"doi":"10.1080/14786440109462720","url":"https://doi.org/10.1080/14786440109462720","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["linear-regression"],"usedBy":["tsne","umap","lora","superposition"],"related":["tsne","umap","lora"],"usedIn":{"zh":"数据可视化、特征压缩、嵌入降维","en":"Visualization, feature compression, embedding reduction"},"resources":[{"title":"scikit-learn 用户指南：矩阵分解（PCA）","url":"https://scikit-learn.org/stable/modules/decomposition.html","type":"docs","lang":"en"},{"title":"An Introduction to Statistical Learning（ISLP，免费）","url":"https://www.statlearning.com/","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/pca/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/pca/"}},{"id":"em","name":{"zh":"期望最大化（EM）","en":"Expectation–Maximization (EM)"},"aka":[],"category":"classic-ml","tier":"extended","year":1977,"difficulty":3,"summary":{"zh":"在有隐藏变量时交替“猜隐藏值—更新参数”。","en":"Alternates inferring hidden variables and updating parameters."},"explain":{"zh":"E 步根据当前参数估计每个样本属于各隐藏状态的概率，M 步用这些概率重新估计参数，似然单调不降。高斯混合模型、HMM 训练都靠它，VAE 的思想也与之相通。","en":"The E-step estimates posterior over hidden variables; the M-step re-fits parameters; likelihood never decreases. It trains GMMs and HMMs and inspires VAEs."},"paper":{"title":"Maximum Likelihood from Incomplete Data Via the EM Algorithm","authors":["Dempster et al."],"venue":"Journal of the Royal Statistical Society Series B: Statistical Methodology","year":1977,"doi":"10.1111/j.2517-6161.1977.tb01600.x","url":"https://doi.org/10.1111/j.2517-6161.1977.tb01600.x","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["kmeans"],"usedBy":["vae"],"related":["vae"],"usedIn":{"zh":"高斯混合模型、语音 HMM、缺失数据","en":"Gaussian mixtures, speech HMMs, missing data"},"resources":[{"title":"scikit-learn 用户指南：高斯混合模型","url":"https://scikit-learn.org/stable/modules/mixture.html","type":"docs","lang":"en"},{"title":"The Elements of Statistical Learning（免费）","url":"https://hastie.su.domains/ElemStatLearn/","type":"book","lang":"en"},{"title":"Bishop：Pattern Recognition and Machine Learning","url":"https://www.microsoft.com/en-us/research/publication/pattern-recognition-machine-learning/","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/em/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/em/"}},{"id":"tsne","name":{"zh":"t-SNE","en":"t-SNE"},"aka":[],"category":"classic-ml","tier":"extended","year":2008,"difficulty":2,"summary":{"zh":"把高维数据压到二维，同时保持“邻居还是邻居”。","en":"Maps high-dimensional data to 2D while keeping neighbors close."},"explain":{"zh":"把高维中的相似度转成概率，在二维里用 t 分布重建同样的邻近关系。非常适合看嵌入空间里的聚类，但簇间距离和大小不可直接解读。","en":"Converts similarities to probabilities and reproduces them in 2D with a t-distribution. Great for seeing clusters in embeddings, but cluster sizes and distances are not meaningful."},"paper":{"title":"Visualizing Data using t-SNE","authors":["van der Maaten & Hinton"],"venue":"JMLR 9","year":2008,"url":"https://jmlr.org/papers/v9/vandermaaten08a.html","source":"官方页面 HTTP/WebFetch 核实","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["pca"],"usedBy":["umap"],"related":["umap"],"usedIn":{"zh":"嵌入可视化、单细胞数据分析","en":"Embedding visualization, single-cell analysis"},"resources":[{"title":"scikit-learn 用户指南：流形学习（t-SNE）","url":"https://scikit-learn.org/stable/modules/manifold.html","type":"docs","lang":"en"},{"title":"Distill：How to Use t-SNE Effectively","url":"https://distill.pub/2016/misread-tsne/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/tsne/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/tsne/"}},{"id":"umap","name":{"zh":"UMAP","en":"UMAP"},"aka":[],"category":"classic-ml","tier":"extended","year":2018,"difficulty":2,"summary":{"zh":"比 t-SNE 更快、更能保留全局结构的降维方法。","en":"A faster dimensionality reduction that better preserves global structure."},"explain":{"zh":"基于流形和拓扑理论构建邻居图，再在低维中优化一个相似的图。速度快、可扩展到百万点，常用于嵌入和单细胞可视化。","en":"Builds a fuzzy neighbor graph from manifold theory and optimizes a similar low-dimensional graph. Fast and scalable to millions of points."},"paper":{"title":"UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction","authors":["Leland McInnes","John Healy","James Melville"],"firstPublished":"2018-02-09","year":2018,"arxiv":"1802.03426","url":"https://arxiv.org/abs/1802.03426","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["pca","tsne"],"usedBy":[],"related":[],"usedIn":{"zh":"嵌入可视化、生物信息","en":"Embedding visualization, bioinformatics"},"resources":[{"title":"UMAP 官方文档","url":"https://umap-learn.readthedocs.io/en/latest/","type":"docs","lang":"en"},{"title":"lmcinnes/umap（官方实现）","url":"https://github.com/lmcinnes/umap","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/umap/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/umap/"}},{"id":"perceptron","name":{"zh":"感知机","en":"Perceptron"},"aka":[],"category":"dl-foundations","tier":"core","year":1958,"difficulty":1,"summary":{"zh":"最早的“人工神经元”：加权求和后过阈值。","en":"The first artificial neuron: weighted sum followed by a threshold."},"explain":{"zh":"每个输入乘一个权重再相加，超过阈值就输出 1；分错了就朝正确方向调整权重。它只能画直线分界，多层堆叠后才有了今天的神经网络。","en":"Multiply inputs by weights, sum, and fire if above threshold; nudge weights when wrong. Alone it draws only linear boundaries — stacking layers gives modern networks."},"paper":{"title":"The perceptron: A probabilistic model for information storage and organization in the brain.","authors":["Rosenblatt et al."],"venue":"Psychological Review","year":1958,"doi":"10.1037/h0042519","url":"https://doi.org/10.1037/h0042519","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":[],"usedBy":["backprop","gelu"],"related":["backprop","logistic-regression"],"usedIn":{"zh":"神经网络的历史起点","en":"Historical starting point of neural networks"},"resources":[{"title":"3Blue1Brown：神经网络系列","url":"https://www.3blue1brown.com/?topic=neural-networks","type":"video","lang":"en"},{"title":"Goodfellow 等：Deep Learning","url":"https://www.deeplearningbook.org/","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/perceptron/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/perceptron/"}},{"id":"sgd","name":{"zh":"随机梯度下降","en":"Stochastic Gradient Descent (SGD)"},"aka":[],"category":"dl-foundations","tier":"core","year":1951,"difficulty":1,"summary":{"zh":"每次用一小批数据估计梯度，沿下坡方向走一小步。","en":"Steps downhill using gradients estimated on small mini-batches."},"explain":{"zh":"损失函数像一片山地，梯度指向上坡，所以往反方向走。每次只用一小批样本估计梯度，又快又能带来有益的噪声；加上动量就是现代优化器的基础。","en":"The loss is a landscape and the gradient points uphill, so step the other way. Mini-batches make it cheap and add helpful noise; momentum turns it into the base of modern optimizers."},"paper":{"title":"A Stochastic Approximation Method","authors":["Robbins et al."],"venue":"The Annals of Mathematical Statistics","year":1951,"doi":"10.1214/aoms/1177729586","url":"https://doi.org/10.1214/aoms/1177729586","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["linear-regression"],"usedBy":["gbdt","backprop","adam","mixed-precision","q-learning","policy-gradient","dp-sgd"],"related":["adam","backprop"],"usedIn":{"zh":"几乎所有神经网络训练","en":"Virtually all neural-network training"},"resources":[{"title":"Sebastian Ruder：An overview of gradient descent optimization algorithms","url":"https://www.ruder.io/optimizing-gradient-descent/","type":"blog","lang":"en"},{"title":"Distill：Why Momentum Really Works","url":"https://distill.pub/2017/momentum/","type":"blog","lang":"en"},{"title":"PyTorch 文档：torch.optim","url":"https://docs.pytorch.org/docs/stable/optim.html","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/sgd/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/sgd/"}},{"id":"backprop","name":{"zh":"反向传播","en":"Backpropagation"},"aka":[],"category":"dl-foundations","tier":"core","year":1986,"difficulty":2,"summary":{"zh":"用链式法则从输出往回算出每个参数的梯度。","en":"Computes every parameter's gradient by applying the chain rule backward."},"explain":{"zh":"前向计算得到损失后，从最后一层开始，把“误差对本层的影响”逐层往回传，一次就算出所有权重的梯度。PyTorch 的自动求导就是它的通用实现。","en":"After the forward pass, propagate the error from the last layer backward with the chain rule, getting all gradients in one sweep. PyTorch autograd is its general implementation."},"paper":{"title":"Learning representations by back-propagating errors","authors":["Rumelhart et al."],"venue":"Nature","year":1986,"doi":"10.1038/323533a0","url":"https://doi.org/10.1038/323533a0","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["sgd","perceptron"],"usedBy":["cnn","lstm","word2vec","dropout","batchnorm","vae","gan","nerf-3dgs","gnn","attribution","grad-cam","adversarial-examples","autoencoder","weight-init"],"related":["adam"],"usedIn":{"zh":"所有深度学习框架的自动求导","en":"Autograd in every deep learning framework"},"resources":[{"title":"colah：Calculus on Computational Graphs: Backpropagation","url":"https://colah.github.io/posts/2015-08-Backprop/","type":"blog","lang":"en"},{"title":"karpathy/micrograd（参考实现）","url":"https://github.com/karpathy/micrograd","type":"repo","lang":"en"},{"title":"Karpathy：Neural Networks: Zero to Hero","url":"https://karpathy.ai/zero-to-hero.html","type":"video","lang":"en"},{"title":"CS231n 笔记：反向传播直觉","url":"https://cs231n.github.io/optimization-2/","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/backprop/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/backprop/"}},{"id":"cnn","name":{"zh":"卷积神经网络","en":"Convolutional Neural Network (CNN)"},"aka":[],"category":"dl-foundations","tier":"core","year":1998,"difficulty":2,"summary":{"zh":"用滑动的小滤波器提取局部图案，层层组合成高级特征。","en":"Slides small learned filters to detect local patterns, composed layer by layer."},"explain":{"zh":"同一个滤波器在整张图上共享参数，先学到边缘、纹理，再组合成部件和物体。LeNet 用于手写识别，AlexNet 在 2012 年引爆深度学习。","en":"Filters share weights across the image, learning edges, then textures, parts and objects. LeNet read digits; AlexNet ignited the deep learning boom in 2012."},"paper":{"title":"Gradient-based learning applied to document recognition","authors":["Lecun et al."],"venue":"Proceedings of the IEEE","year":1998,"doi":"10.1109/5.726791","url":"https://doi.org/10.1109/5.726791","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["backprop"],"usedBy":["resnet","vit","dqn","mcts-alphazero","grad-cam","adversarial-examples","unet","faster-rcnn","yolo"],"related":["resnet","vit"],"usedIn":{"zh":"图像分类、检测、医学影像","en":"Image classification, detection, medical imaging"},"resources":[{"title":"Stanford CS231n 视觉识别中的卷积网络","url":"https://cs231n.github.io/","type":"course","lang":"en"},{"title":"Dive into Deep Learning（D2L）","url":"https://d2l.ai/","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/cnn/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/cnn/"}},{"id":"lstm","name":{"zh":"LSTM 长短期记忆网络","en":"Long Short-Term Memory (LSTM)"},"aka":[],"category":"dl-foundations","tier":"core","year":1997,"difficulty":2,"summary":{"zh":"带“门”的循环网络，能记住长距离信息。","en":"A gated recurrent network that preserves long-range information."},"explain":{"zh":"普通 RNN 的梯度会随时间指数衰减。LSTM 用一条细胞状态“传送带”加输入门、遗忘门、输出门控制信息进出，解决了长期依赖。Transformer 出现前它统治序列建模，xLSTM 让它重返舞台。","en":"Plain RNN gradients vanish over time. LSTM adds a cell-state conveyor with input/forget/output gates. It ruled sequence modeling before Transformers; xLSTM revisits it."},"paper":{"title":"Long Short-Term Memory","authors":["Hochreiter et al."],"venue":"Neural Computation","year":1997,"doi":"10.1162/neco.1997.9.8.1735","url":"https://doi.org/10.1162/neco.1997.9.8.1735","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["backprop"],"usedBy":["seq2seq","s4","rwkv","xlstm","world-models","gru","elmo","ctc"],"related":["seq2seq","xlstm","transformer"],"usedIn":{"zh":"语音识别、机器翻译（2014–2017）","en":"Speech recognition, machine translation (2014–2017)"},"resources":[{"title":"colah：Understanding LSTM Networks","url":"https://colah.github.io/posts/2015-08-Understanding-LSTMs/","type":"blog","lang":"en"},{"title":"Dive into Deep Learning（D2L）","url":"https://d2l.ai/","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/lstm/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/lstm/"}},{"id":"seq2seq","name":{"zh":"序列到序列（Seq2Seq）","en":"Sequence-to-Sequence Learning"},"aka":[],"category":"dl-foundations","tier":"core","year":2014,"difficulty":2,"summary":{"zh":"一个网络读入句子，另一个网络生成句子。","en":"An encoder reads a sequence and a decoder generates another."},"explain":{"zh":"编码器把输入压成向量，解码器再逐词生成输出，端到端训练机器翻译。它暴露的“信息瓶颈”问题催生了注意力机制。","en":"An encoder compresses the input and a decoder emits the output word by word, trained end to end. Its bottleneck motivated attention."},"paper":{"title":"Sequence to Sequence Learning with Neural Networks","authors":["Ilya Sutskever","Oriol Vinyals","Quoc V. Le"],"firstPublished":"2014-09-10","year":2014,"arxiv":"1409.3215","url":"https://arxiv.org/abs/1409.3215","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["lstm"],"usedBy":["bahdanau-attention"],"related":["bahdanau-attention","transformer"],"usedIn":{"zh":"早期神经机器翻译","en":"Early neural machine translation"},"resources":[{"title":"Jay Alammar：Visualizing Seq2seq Models with Attention","url":"https://jalammar.github.io/visualizing-neural-machine-translation-mechanics-of-seq2seq-models-with-attention/","type":"blog","lang":"en"},{"title":"Dive into Deep Learning（D2L）","url":"https://d2l.ai/","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/seq2seq/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/seq2seq/"}},{"id":"word2vec","name":{"zh":"Word2Vec 词向量","en":"Word2Vec"},"aka":[],"category":"dl-foundations","tier":"core","year":2013,"difficulty":1,"summary":{"zh":"用“上下文预测词”学到能做加减法的词向量。","en":"Learns word vectors by predicting words from context."},"explain":{"zh":"把每个词映射成稠密向量，让出现在相似上下文的词彼此靠近，于是出现“国王−男人+女人≈女王”的规律。它是 embedding 思想的普及者。","en":"Maps each word to a dense vector so words in similar contexts are close, giving king − man + woman ≈ queen. It popularized embeddings."},"paper":{"title":"Efficient Estimation of Word Representations in Vector Space","authors":["Tomas Mikolov","Kai Chen","Greg Corrado","et al."],"firstPublished":"2013-01-16","year":2013,"arxiv":"1301.3781","url":"https://arxiv.org/abs/1301.3781","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["backprop"],"usedBy":["glove","elmo"],"related":["bpe","sentence-bert"],"usedIn":{"zh":"词嵌入、推荐系统 item2vec","en":"Word embeddings, item2vec recommenders"},"resources":[{"title":"Jay Alammar：The Illustrated Word2vec","url":"https://jalammar.github.io/illustrated-word2vec/","type":"blog","lang":"en"},{"title":"Stanford CS224N 深度学习自然语言处理","url":"https://web.stanford.edu/class/cs224n/","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/word2vec/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/word2vec/"}},{"id":"adam","name":{"zh":"Adam 优化器","en":"Adam"},"aka":[],"category":"dl-foundations","tier":"core","year":2014,"difficulty":2,"summary":{"zh":"给每个参数自适应步长的动量优化器。","en":"A momentum optimizer with per-parameter adaptive step sizes."},"explain":{"zh":"同时记录梯度的滑动平均（动量）和梯度平方的滑动平均，用后者给每个参数单独缩放步长。几乎是深度学习的默认优化器。","en":"Tracks running means of gradients (momentum) and squared gradients, scaling each parameter's step. It is the default deep learning optimizer."},"paper":{"title":"Adam: A Method for Stochastic Optimization","authors":["Diederik P. Kingma","Jimmy Ba"],"firstPublished":"2014-12-22","year":2014,"arxiv":"1412.6980","url":"https://arxiv.org/abs/1412.6980","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["sgd"],"usedBy":["adamw"],"related":["adamw","muon"],"usedIn":{"zh":"几乎所有神经网络训练","en":"Nearly all neural-network training"},"resources":[{"title":"Sebastian Ruder：An overview of gradient descent optimization algorithms","url":"https://www.ruder.io/optimizing-gradient-descent/","type":"blog","lang":"en"},{"title":"PyTorch 文档：torch.optim","url":"https://docs.pytorch.org/docs/stable/optim.html","type":"docs","lang":"en"},{"title":"Dive into Deep Learning（D2L）","url":"https://d2l.ai/","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/adam/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/adam/"}},{"id":"adamw","name":{"zh":"AdamW","en":"AdamW (Decoupled Weight Decay)"},"aka":[],"category":"dl-foundations","tier":"core","year":2017,"difficulty":2,"summary":{"zh":"把权重衰减从 Adam 的梯度更新中解耦出来。","en":"Decouples weight decay from Adam's gradient update."},"explain":{"zh":"在 Adam 里把 L2 正则加到梯度上，效果会被自适应步长扭曲；AdamW 直接按比例缩小权重，正则更有效。它是训练大语言模型的标准优化器。","en":"Adding L2 to Adam's gradient gets distorted by adaptive scaling; AdamW shrinks weights directly. It is the standard optimizer for LLM training."},"paper":{"title":"Decoupled Weight Decay Regularization","authors":["Ilya Loshchilov","Frank Hutter"],"firstPublished":"2017-11-14","year":2017,"arxiv":"1711.05101","url":"https://arxiv.org/abs/1711.05101","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["adam"],"usedBy":["muon","mup"],"related":["muon"],"usedIn":{"zh":"GPT、Llama 等大模型预训练","en":"Pretraining of GPT, Llama and most LLMs"},"resources":[{"title":"PyTorch 文档：torch.optim","url":"https://docs.pytorch.org/docs/stable/optim.html","type":"docs","lang":"en"},{"title":"Hugging Face：Smol Training Playbook","url":"https://huggingface.co/spaces/HuggingFaceTB/smol-training-playbook","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/adamw/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/adamw/"}},{"id":"muon","name":{"zh":"Muon 优化器","en":"Muon Optimizer"},"aka":[],"category":"dl-foundations","tier":"core","year":2024,"difficulty":4,"summary":{"zh":"把动量更新“正交化”后再用于矩阵参数的新优化器。","en":"An optimizer that orthogonalizes momentum updates for matrix parameters."},"explain":{"zh":"对二维权重矩阵，先算动量，再用 Newton–Schulz 迭代把它近似变成正交矩阵，使各方向更新幅度均衡。2024 年由 Keller Jordan 在 nanoGPT 速通中提出，Moonshot 在 2025 年验证其可扩展到大模型。","en":"For weight matrices, take the momentum and approximately orthogonalize it with Newton–Schulz iterations so all directions update evenly. Proposed by Keller Jordan in 2024 nanoGPT speedruns; Moonshot showed it scales in 2025."},"paper":{"title":"Muon: An optimizer for hidden layers in neural networks (blog)","authors":["Keller Jordan"],"venue":"Blog","year":2024,"url":"https://kellerjordan.github.io/posts/muon/","source":"官方页面 HTTP/WebFetch 核实","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["adamw"],"usedBy":[],"related":["adamw"],"usedIn":{"zh":"Moonshot Kimi 系列大模型训练（MuonClip）；i14 Laplace 3B 用 Muon 训练 2D 权重、AdamW 训练嵌入与归一化（AWS 博客 2026-10-20）；TorchTitan 支持分布式 Muon","en":"Moonshot Kimi training (MuonClip); i14 Laplace 3B trains 2D weights with Muon and embeddings/norms with AdamW (AWS blog, Oct 2026); TorchTitan supports distributed Muon"},"resources":[{"title":"Keller Jordan：Muon 博客","url":"https://kellerjordan.github.io/posts/muon/","type":"blog","lang":"en"},{"title":"KellerJordan/Muon（官方实现）","url":"https://github.com/KellerJordan/Muon","type":"repo","lang":"en"},{"title":"AWS 机器学习博客：在 Trainium2 上训练与服务 i14 Laplace 3B 线性注意力模型（2026-10-20）","url":"https://aws.amazon.com/blogs/machine-learning/beyond-the-kv-cache-training-and-serving-i14s-laplace-3b-linear-attention-model-on-aws-trainium2-with-native-pytorch/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/muon/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/muon/"}},{"id":"dropout","name":{"zh":"Dropout","en":"Dropout"},"aka":[],"category":"dl-foundations","tier":"core","year":2012,"difficulty":1,"summary":{"zh":"训练时随机“关掉”一部分神经元以防过拟合。","en":"Randomly drops neurons during training to prevent overfitting."},"explain":{"zh":"每一步随机让一部分神经元输出为 0，网络无法依赖任何单个神经元，相当于同时训练很多子网络取平均。","en":"Each step zeroes random units so the network cannot rely on any single one — like averaging many sub-networks."},"paper":{"title":"Improving neural networks by preventing co-adaptation of feature detectors","authors":["Geoffrey E. Hinton","Nitish Srivastava","Alex Krizhevsky","et al."],"firstPublished":"2012-07-03","year":2012,"arxiv":"1207.0580","url":"https://arxiv.org/abs/1207.0580","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["backprop"],"usedBy":[],"related":["batchnorm"],"usedIn":{"zh":"CNN、全连接网络、部分 Transformer","en":"CNNs, MLPs, some Transformers"},"resources":[{"title":"Dive into Deep Learning（D2L）","url":"https://d2l.ai/","type":"book","lang":"en"},{"title":"Stanford CS231n 视觉识别中的卷积网络","url":"https://cs231n.github.io/","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/dropout/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/dropout/"}},{"id":"batchnorm","name":{"zh":"批归一化","en":"Batch Normalization"},"aka":[],"category":"dl-foundations","tier":"core","year":2015,"difficulty":2,"summary":{"zh":"把每层输出按一个 batch 标准化，让深网络更好训练。","en":"Normalizes activations across a batch to stabilize deep training."},"explain":{"zh":"对每个通道减去 batch 均值、除以标准差，再学一个缩放和偏移。训练更稳、可用更大学习率，是 CNN 时代的标配。","en":"Subtract the batch mean, divide by std per channel, then learn a scale and shift. Enables stable, faster training in CNNs."},"paper":{"title":"Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift","authors":["Sergey Ioffe","Christian Szegedy"],"firstPublished":"2015-02-11","year":2015,"arxiv":"1502.03167","url":"https://arxiv.org/abs/1502.03167","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["backprop"],"usedBy":["layernorm","resnet"],"related":["layernorm","rmsnorm"],"usedIn":{"zh":"ResNet 等 CNN","en":"ResNet and most CNNs"},"resources":[{"title":"Dive into Deep Learning（D2L）","url":"https://d2l.ai/","type":"book","lang":"en"},{"title":"Stanford CS231n 视觉识别中的卷积网络","url":"https://cs231n.github.io/","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/batchnorm/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/batchnorm/"}},{"id":"layernorm","name":{"zh":"层归一化","en":"Layer Normalization"},"aka":[],"category":"dl-foundations","tier":"core","year":2016,"difficulty":2,"summary":{"zh":"在单个样本内部做标准化，不依赖 batch。","en":"Normalizes within each example instead of across the batch."},"explain":{"zh":"对一个 token 的所有特征维度求均值方差再标准化，适合变长序列和小 batch。Transformer 默认使用它。","en":"Normalizes across a token's features, so it works with variable-length sequences and small batches. The Transformer default."},"paper":{"title":"Layer Normalization","authors":["Jimmy Lei Ba","Jamie Ryan Kiros","Geoffrey E. Hinton"],"firstPublished":"2016-07-21","year":2016,"arxiv":"1607.06450","url":"https://arxiv.org/abs/1607.06450","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["batchnorm"],"usedBy":["rmsnorm","transformer"],"related":["rmsnorm","transformer"],"usedIn":{"zh":"BERT、GPT-2 等 Transformer","en":"BERT, GPT-2 and other Transformers"},"resources":[{"title":"Dive into Deep Learning（D2L）","url":"https://d2l.ai/","type":"book","lang":"en"},{"title":"PyTorch 文档：torch.nn.LayerNorm","url":"https://docs.pytorch.org/docs/stable/generated/torch.nn.LayerNorm.html","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/layernorm/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/layernorm/"}},{"id":"rmsnorm","name":{"zh":"RMSNorm","en":"RMSNorm"},"aka":[],"category":"dl-foundations","tier":"core","year":2019,"difficulty":2,"summary":{"zh":"只用均方根缩放、省掉减均值的简化版 LayerNorm。","en":"A simplified LayerNorm that only rescales by root-mean-square."},"explain":{"zh":"去掉“减均值”和偏置，只除以均方根，计算更省而效果相当。现代开源大模型几乎都用它。","en":"Drops mean-centering and bias, only dividing by RMS — cheaper with similar quality. Used by nearly all modern open LLMs."},"paper":{"title":"Root Mean Square Layer Normalization","authors":["Biao Zhang","Rico Sennrich"],"firstPublished":"2019-10-16","year":2019,"arxiv":"1910.07467","url":"https://arxiv.org/abs/1910.07467","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["layernorm"],"usedBy":[],"related":["swiglu"],"usedIn":{"zh":"Llama、Qwen、DeepSeek 等","en":"Llama, Qwen, DeepSeek and others"},"resources":[{"title":"rasbt/LLMs-from-scratch","url":"https://github.com/rasbt/LLMs-from-scratch","type":"repo","lang":"en"},{"title":"PyTorch 文档：torch.nn.RMSNorm","url":"https://docs.pytorch.org/docs/stable/generated/torch.nn.RMSNorm.html","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/rmsnorm/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/rmsnorm/"}},{"id":"resnet","name":{"zh":"残差网络","en":"Residual Networks (ResNet)"},"aka":[],"category":"dl-foundations","tier":"core","year":2015,"difficulty":2,"summary":{"zh":"加一条“捷径”让信号绕过层，使上百层网络可训练。","en":"Skip connections let signals bypass layers, making very deep nets trainable."},"explain":{"zh":"每层只学习“在输入基础上的增量”，输出 = 输入 + F(输入)。梯度能沿捷径直通，深度从十几层跳到上百层。Transformer 的每个子层也都套着残差连接。","en":"Each block learns a residual: output = input + F(input), so gradients flow through the shortcut. Depth jumped to 100+ layers; every Transformer sublayer uses it."},"paper":{"title":"Deep Residual Learning for Image Recognition","authors":["Kaiming He","Xiangyu Zhang","Shaoqing Ren","et al."],"firstPublished":"2015-12-10","year":2015,"arxiv":"1512.03385","url":"https://arxiv.org/abs/1512.03385","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["cnn","batchnorm"],"usedBy":["transformer","simclr","faster-rcnn"],"related":["transformer"],"usedIn":{"zh":"视觉骨干网络；Transformer 中的残差连接","en":"Vision backbones; residual paths in Transformers"},"resources":[{"title":"Dive into Deep Learning（D2L）","url":"https://d2l.ai/","type":"book","lang":"en"},{"title":"Stanford CS231n 视觉识别中的卷积网络","url":"https://cs231n.github.io/","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/resnet/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/resnet/"}},{"id":"gelu","name":{"zh":"GELU 激活函数","en":"GELU"},"aka":[],"category":"dl-foundations","tier":"core","year":2016,"difficulty":2,"summary":{"zh":"按概率“软开关”输入的平滑激活函数。","en":"A smooth activation that gates inputs by their Gaussian probability."},"explain":{"zh":"把输入乘以它小于标准正态随机变量的概率，比 ReLU 更平滑。BERT、GPT 系列早期都用它。","en":"Multiplies the input by the standard normal CDF, a smoother ReLU. Used in BERT and GPT-family models."},"paper":{"title":"Gaussian Error Linear Units (GELUs)","authors":["Dan Hendrycks","Kevin Gimpel"],"firstPublished":"2016-06-27","year":2016,"arxiv":"1606.08415","url":"https://arxiv.org/abs/1606.08415","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["perceptron"],"usedBy":["swiglu"],"related":["swiglu"],"usedIn":{"zh":"BERT、GPT-2/3、ViT","en":"BERT, GPT-2/3, ViT"},"resources":[{"title":"Dive into Deep Learning（D2L）","url":"https://d2l.ai/","type":"book","lang":"en"},{"title":"PyTorch 文档：torch.nn.GELU","url":"https://docs.pytorch.org/docs/stable/generated/torch.nn.GELU.html","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/gelu/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/gelu/"}},{"id":"swiglu","name":{"zh":"SwiGLU","en":"SwiGLU (Gated Linear Units)"},"aka":[],"category":"dl-foundations","tier":"core","year":2020,"difficulty":3,"summary":{"zh":"给前馈层加一条“门控”支路，提升 Transformer 质量。","en":"Adds a gating branch to the feed-forward layer to improve quality."},"explain":{"zh":"前馈层把输入投影成两份，一份过 Swish 激活当“门”，与另一份逐元素相乘。参数量相同下效果更好，成为现代大模型标配。","en":"The FFN projects the input twice; one goes through Swish as a gate multiplying the other. Better at equal parameters and now standard."},"paper":{"title":"GLU Variants Improve Transformer","authors":["Noam Shazeer"],"firstPublished":"2020-02-12","year":2020,"arxiv":"2002.05202","url":"https://arxiv.org/abs/2002.05202","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["gelu","transformer"],"usedBy":[],"related":["rmsnorm"],"usedIn":{"zh":"PaLM、Llama、Qwen 等","en":"PaLM, Llama, Qwen and others"},"resources":[{"title":"rasbt/LLMs-from-scratch","url":"https://github.com/rasbt/LLMs-from-scratch","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/swiglu/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/swiglu/"}},{"id":"bpe","name":{"zh":"BPE 子词分词","en":"Byte-Pair Encoding (BPE) Tokenization"},"aka":[],"category":"transformer","tier":"core","year":2015,"difficulty":1,"summary":{"zh":"反复合并最常见的字符对，得到子词词表。","en":"Builds a subword vocabulary by repeatedly merging the most frequent pairs."},"explain":{"zh":"从单个字符（或字节）开始，统计哪两个相邻符号最常一起出现就合并成新 token，直到词表够大。这样既能表示任意新词，又不会让序列太长；GPT、Llama 的分词器都基于它。","en":"Start from characters or bytes and repeatedly merge the most frequent adjacent pair into a new token. It can spell any word without overly long sequences; GPT and Llama tokenizers build on it."},"paper":{"title":"Neural Machine Translation of Rare Words with Subword Units","authors":["Rico Sennrich","Barry Haddow","Alexandra Birch"],"firstPublished":"2015-08-31","year":2015,"arxiv":"1508.07909","url":"https://arxiv.org/abs/1508.07909","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":[],"usedBy":["sentencepiece"],"related":["word2vec","transformer"],"usedIn":{"zh":"GPT 系列 tiktoken、Llama、Qwen 分词器","en":"Tokenizers of GPT (tiktoken), Llama, Qwen"},"resources":[{"title":"karpathy/minbpe","url":"https://github.com/karpathy/minbpe","type":"repo","lang":"en"},{"title":"HF LLM 课程：BPE 分词","url":"https://huggingface.co/learn/llm-course/chapter6/5","type":"course","lang":"en"},{"title":"openai/tiktoken","url":"https://github.com/openai/tiktoken","type":"repo","lang":"en"},{"title":"google/sentencepiece","url":"https://github.com/google/sentencepiece","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/bpe/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/bpe/"}},{"id":"bahdanau-attention","name":{"zh":"注意力机制（Bahdanau）","en":"Attention (Bahdanau)"},"aka":[],"category":"transformer","tier":"core","year":2014,"difficulty":2,"summary":{"zh":"生成每个词时“回头看”输入里最相关的部分。","en":"Lets the decoder look back at the most relevant input positions."},"explain":{"zh":"解码器每一步对编码器所有位置打分，按分数加权求和得到上下文向量，不再把整句压成一个向量。这就是后来 Transformer 注意力的前身。","en":"At each step the decoder scores all encoder positions and takes a weighted sum, instead of squeezing the sentence into one vector. It is the precursor of Transformer attention."},"paper":{"title":"Neural Machine Translation by Jointly Learning to Align and Translate","authors":["Dzmitry Bahdanau","Kyunghyun Cho","Yoshua Bengio"],"firstPublished":"2014-09-01","year":2014,"arxiv":"1409.0473","url":"https://arxiv.org/abs/1409.0473","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["seq2seq"],"usedBy":["transformer"],"related":["transformer"],"usedIn":{"zh":"神经机器翻译","en":"Neural machine translation"},"resources":[{"title":"Jay Alammar：Visualizing Seq2seq Models with Attention","url":"https://jalammar.github.io/visualizing-neural-machine-translation-mechanics-of-seq2seq-models-with-attention/","type":"blog","lang":"en"},{"title":"Lilian Weng：Attention? Attention!","url":"https://lilianweng.github.io/posts/2018-06-24-attention/","type":"blog","lang":"en"},{"title":"Distill：Attention and Augmented RNNs","url":"https://distill.pub/2016/augmented-rnns/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/bahdanau-attention/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/bahdanau-attention/"}},{"id":"transformer","name":{"zh":"Transformer 与自注意力","en":"Transformer & Self-Attention"},"aka":[],"category":"transformer","tier":"core","year":2017,"difficulty":2,"summary":{"zh":"完全基于注意力、可并行训练的序列模型，是所有大模型的骨架。","en":"An attention-only, parallelizable sequence model — the backbone of all LLMs."},"explain":{"zh":"每个 token 生成查询（Q）、键（K）、值（V），用 Q 与所有 K 的相似度给 V 加权求和，从而直接“看到”整句。多头注意力 + 前馈层 + 残差 + 归一化堆叠成层，训练可完全并行。","en":"Each token emits a query, key and value; softmax(QKᵀ)·V lets every token attend to all others. Multi-head attention, FFN, residuals and normalization stack into layers trained fully in parallel."},"paper":{"title":"Attention Is All You Need","authors":["Ashish Vaswani","Noam Shazeer","Niki Parmar","et al."],"firstPublished":"2017-06-12","year":2017,"arxiv":"1706.03762","url":"https://arxiv.org/abs/1706.03762","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["bahdanau-attention","resnet","layernorm"],"usedBy":["swiglu","bert","t5","gpt3-icl","rope","alibi","mqa","flashattention","sliding-window","linear-attention","sparse-moe","mtp","tensor-parallel","kv-cache","sampling","var","vit","flamingo","whisper","genie","alphafold","alphageometry","logit-lens","induction-heads","rome","looped-lm","detr","decision-transformer","wav2vec2"],"related":["bert","gpt3-icl","flashattention","rope"],"usedIn":{"zh":"GPT、Claude、Gemini、Llama、Qwen、DeepSeek 等全部大模型","en":"Every LLM: GPT, Claude, Gemini, Llama, Qwen, DeepSeek"},"resources":[{"title":"Jay Alammar：The Illustrated Transformer","url":"https://jalammar.github.io/illustrated-transformer/","type":"blog","lang":"en"},{"title":"Harvard NLP：The Annotated Transformer","url":"https://nlp.seas.harvard.edu/annotated-transformer/","type":"blog","lang":"en"},{"title":"Karpathy：Neural Networks: Zero to Hero","url":"https://karpathy.ai/zero-to-hero.html","type":"video","lang":"en"},{"title":"karpathy/nanoGPT（参考实现）","url":"https://github.com/karpathy/nanoGPT","type":"repo","lang":"en"},{"title":"Stanford CS224N 深度学习自然语言处理","url":"https://web.stanford.edu/class/cs224n/","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/transformer/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/transformer/"}},{"id":"bert","name":{"zh":"BERT","en":"BERT"},"aka":[],"category":"transformer","tier":"core","year":2018,"difficulty":2,"summary":{"zh":"双向 Transformer 编码器，通过“完形填空”预训练。","en":"A bidirectional Transformer encoder pretrained by masked-word prediction."},"explain":{"zh":"随机遮住 15% 的词让模型去猜，模型同时看左右上下文，学到很强的语义表示。它开启了“预训练+微调”时代，至今仍是嵌入和分类模型的骨干。","en":"Mask ~15% of tokens and predict them using both left and right context. It started the pretrain-then-finetune era and still underlies embedding and classification models."},"paper":{"title":"BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding","authors":["Jacob Devlin","Ming-Wei Chang","Kenton Lee","et al."],"firstPublished":"2018-10-11","year":2018,"arxiv":"1810.04805","url":"https://arxiv.org/abs/1810.04805","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["transformer"],"usedBy":["t5","diffusion-lm","adapters","sentence-bert","colbert","mae","cross-encoder-rerank"],"related":["sentence-bert","t5","colbert"],"usedIn":{"zh":"搜索相关性、文本分类、嵌入模型","en":"Search relevance, classification, embedding models"},"resources":[{"title":"Jay Alammar：The Illustrated BERT","url":"https://jalammar.github.io/illustrated-bert/","type":"blog","lang":"en"},{"title":"Hugging Face LLM 课程","url":"https://huggingface.co/learn/llm-course/chapter1/1","type":"course","lang":"multi"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/bert/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/bert/"}},{"id":"t5","name":{"zh":"T5 文本到文本","en":"T5 (Text-to-Text Transfer Transformer)"},"aka":[],"category":"transformer","tier":"core","year":2019,"difficulty":2,"summary":{"zh":"把所有 NLP 任务统一成“输入文本→输出文本”。","en":"Casts every NLP task as text in, text out."},"explain":{"zh":"翻译、摘要、分类都写成文本前缀，用同一个编码器-解码器模型训练。论文系统比较了预训练目标、数据和规模，是重要的实验指南。","en":"Translation, summarization and classification all become prefixed text tasks for one encoder–decoder. The paper is a systematic study of objectives, data and scale."},"paper":{"title":"Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer","authors":["Colin Raffel","Noam Shazeer","Adam Roberts","et al."],"firstPublished":"2019-10-23","year":2019,"arxiv":"1910.10683","url":"https://arxiv.org/abs/1910.10683","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["transformer","bert"],"usedBy":["instruction-tuning"],"related":["instruction-tuning"],"usedIn":{"zh":"Flan-T5、文本生成编码器","en":"Flan-T5; text encoders for image models"},"resources":[{"title":"Hugging Face LLM 课程","url":"https://huggingface.co/learn/llm-course/chapter1/1","type":"course","lang":"multi"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/t5/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/t5/"}},{"id":"gpt3-icl","name":{"zh":"GPT-3 与上下文学习","en":"GPT-3 & In-Context Learning"},"aka":[],"category":"transformer","tier":"core","year":2020,"difficulty":2,"summary":{"zh":"足够大的语言模型只看几个例子就能完成新任务。","en":"Large enough language models learn new tasks from a few examples in the prompt."},"explain":{"zh":"GPT-3 只做“预测下一个词”的预训练，却能在提示里给几个示例后直接完成翻译、问答等任务，无需更新参数。这让“提示词”成为一种新的编程方式。","en":"Trained only to predict the next token, GPT-3 performs new tasks from a few in-prompt examples without weight updates — making prompting a new kind of programming."},"paper":{"title":"Language Models are Few-Shot Learners","authors":["Tom B. Brown","Benjamin Mann","Nick Ryder","et al."],"firstPublished":"2020-05-28","year":2020,"arxiv":"2005.14165","url":"https://arxiv.org/abs/2005.14165","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["transformer"],"usedBy":["scaling-laws","cot","instruction-tuning","rag","toolformer","valle"],"related":["scaling-laws","cot"],"usedIn":{"zh":"所有通用聊天模型的提示用法","en":"Prompting of every general-purpose chat model"},"resources":[{"title":"Jay Alammar：The Illustrated GPT-2","url":"https://jalammar.github.io/illustrated-gpt2/","type":"blog","lang":"en"},{"title":"Lilian Weng：Prompt Engineering","url":"https://lilianweng.github.io/posts/2023-03-15-prompt-engineering/","type":"blog","lang":"en"},{"title":"karpathy/nanoGPT（参考实现）","url":"https://github.com/karpathy/nanoGPT","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/gpt3-icl/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/gpt3-icl/"}},{"id":"rope","name":{"zh":"旋转位置编码 RoPE","en":"Rotary Position Embedding (RoPE)"},"aka":[],"category":"transformer","tier":"core","year":2021,"difficulty":3,"summary":{"zh":"通过旋转 Q/K 向量来编码位置，使注意力天然感知相对距离。","en":"Encodes position by rotating query/key vectors so attention sees relative distance."},"explain":{"zh":"把向量按维度两两成对，在复平面上按位置旋转不同角度；两个 token 的点积只依赖它们的相对距离。它是现代大模型的标配，也是长上下文扩展（YaRN 等）的基础。","en":"Pair up dimensions and rotate them by position-dependent angles; the dot product then depends only on relative offset. Standard in modern LLMs and the base for context extension like YaRN."},"paper":{"title":"RoFormer: Enhanced Transformer with Rotary Position Embedding","authors":["Jianlin Su","Yu Lu","Shengfeng Pan","et al."],"firstPublished":"2021-04-20","year":2021,"arxiv":"2104.09864","url":"https://arxiv.org/abs/2104.09864","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["transformer"],"usedBy":["yarn"],"related":["alibi","yarn"],"usedIn":{"zh":"Llama、Qwen、DeepSeek、Mistral 等","en":"Llama, Qwen, DeepSeek, Mistral and more"},"resources":[{"title":"EleutherAI 博客：Rotary Embeddings","url":"https://blog.eleuther.ai/rotary-embeddings/","type":"blog","lang":"en"},{"title":"Hugging Face 博客：You could have designed state of the art positional encoding","url":"https://huggingface.co/blog/designing-positional-encoding","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/rope/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/rope/"}},{"id":"alibi","name":{"zh":"ALiBi 线性偏置","en":"ALiBi (Attention with Linear Biases)"},"aka":[],"category":"transformer","tier":"extended","year":2021,"difficulty":3,"summary":{"zh":"按距离给注意力分数加线性惩罚，代替位置编码。","en":"Replaces position embeddings with a distance-proportional attention penalty."},"explain":{"zh":"不加位置向量，而是让越远的 token 注意力分数扣得越多。训练短、推理长时外推更好。","en":"No position vectors — farther tokens just get a larger penalty on their attention scores, which extrapolates to longer inputs."},"paper":{"title":"Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation","authors":["Ofir Press","Noah A. Smith","Mike Lewis"],"firstPublished":"2021-08-27","year":2021,"arxiv":"2108.12409","url":"https://arxiv.org/abs/2108.12409","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["transformer"],"usedBy":[],"related":["rope"],"usedIn":{"zh":"BLOOM、MPT","en":"BLOOM, MPT"},"resources":[{"title":"Lilian Weng：The Transformer Family v2","url":"https://lilianweng.github.io/posts/2023-01-27-the-transformer-family-v2/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/alibi/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/alibi/"}},{"id":"mqa","name":{"zh":"多查询注意力 MQA","en":"Multi-Query Attention (MQA)"},"aka":[],"category":"transformer","tier":"extended","year":2019,"difficulty":3,"summary":{"zh":"所有注意力头共享一份 K/V，大幅缩小 KV Cache。","en":"All heads share one key/value head, shrinking the KV cache."},"explain":{"zh":"标准多头注意力每个头都有自己的 K、V；MQA 只保留一份，解码时读取的显存少很多，速度更快，代价是少量质量下降。","en":"Instead of per-head keys and values, keep one shared pair. Decoding reads far less memory and is faster, with a small quality cost."},"paper":{"title":"Fast Transformer Decoding: One Write-Head is All You Need","authors":["Noam Shazeer"],"firstPublished":"2019-11-06","year":2019,"arxiv":"1911.02150","url":"https://arxiv.org/abs/1911.02150","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["transformer","kv-cache"],"usedBy":["gqa"],"related":["gqa","mla"],"usedIn":{"zh":"PaLM、Falcon","en":"PaLM, Falcon"},"resources":[{"title":"Lilian Weng：Large Transformer Model Inference Optimization","url":"https://lilianweng.github.io/posts/2023-01-10-inference-optimization/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/mqa/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/mqa/"}},{"id":"gqa","name":{"zh":"分组查询注意力 GQA","en":"Grouped-Query Attention (GQA)"},"aka":[],"category":"transformer","tier":"core","year":2023,"difficulty":3,"summary":{"zh":"把注意力头分组，每组共享一份 K/V，在 MHA 与 MQA 间折中。","en":"Groups heads to share K/V, a middle ground between MHA and MQA."},"explain":{"zh":"例如 32 个查询头分成 8 组，每组共享一份 K/V，KV Cache 缩小 4 倍而质量几乎不降。它是当前开源模型最常用的注意力形式。","en":"E.g., 32 query heads share 8 KV heads, cutting the KV cache 4× with almost no quality loss. The most common attention variant in open models."},"paper":{"title":"GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints","authors":["Joshua Ainslie","James Lee-Thorp","Michiel de Jong","et al."],"firstPublished":"2023-05-22","year":2023,"arxiv":"2305.13245","url":"https://arxiv.org/abs/2305.13245","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["mqa"],"usedBy":["mla"],"related":["mla","kv-cache"],"usedIn":{"zh":"Llama 2 70B、Llama 3、Mistral、Qwen 等","en":"Llama 2 70B, Llama 3, Mistral, Qwen"},"resources":[{"title":"rasbt/LLMs-from-scratch","url":"https://github.com/rasbt/LLMs-from-scratch","type":"repo","lang":"en"},{"title":"Transformers 文档：KV Cache","url":"https://huggingface.co/docs/transformers/main/en/kv_cache","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/gqa/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/gqa/"}},{"id":"mla","name":{"zh":"多头潜在注意力 MLA","en":"Multi-head Latent Attention (MLA)"},"aka":[],"category":"transformer","tier":"core","year":2024,"difficulty":4,"summary":{"zh":"把 K/V 压缩成一个低维潜在向量缓存，再按需还原。","en":"Caches a compressed low-rank latent instead of full keys and values."},"explain":{"zh":"每个 token 只缓存一个小的潜在向量，计算注意力时再投影回各头的 K/V，KV Cache 远小于 GQA 而效果接近标准多头注意力。DeepSeek-V2 提出，V3/R1 沿用。","en":"Each token caches one small latent vector that is projected back to per-head keys/values at attention time — much smaller than GQA with MHA-like quality. Introduced in DeepSeek-V2, kept in V3/R1."},"paper":{"title":"DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model","authors":[" DeepSeek-AI","Aixin Liu","Bei Feng","et al."],"firstPublished":"2024-05-07","year":2024,"arxiv":"2405.04434","url":"https://arxiv.org/abs/2405.04434","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["gqa","kv-cache"],"usedBy":[],"related":["deepseek-moe"],"usedIn":{"zh":"DeepSeek-V2/V3/R1、Kimi K2","en":"DeepSeek-V2/V3/R1, Kimi K2"},"resources":[{"title":"deepseek-ai/FlashMLA","url":"https://github.com/deepseek-ai/FlashMLA","type":"repo","lang":"en"},{"title":"deepseek-ai/DeepSeek-V3","url":"https://github.com/deepseek-ai/DeepSeek-V3","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/mla/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/mla/"}},{"id":"flashattention","name":{"zh":"FlashAttention","en":"FlashAttention (v1–v3)"},"aka":[],"category":"transformer","tier":"core","year":2022,"difficulty":4,"summary":{"zh":"按 GPU 内存层级分块计算的精确注意力，快且省显存。","en":"Exact attention computed in GPU-memory-aware tiles — faster and lighter."},"explain":{"zh":"注意力的瓶颈不是算力而是显存读写。FlashAttention 把 Q/K/V 切块放进片上高速缓存，边算边做在线 softmax，从不把完整注意力矩阵写回显存。v2 改进并行，v3 利用 Hopper GPU 异步与 FP8。","en":"Attention is memory-bound, not compute-bound. FlashAttention tiles Q/K/V into on-chip SRAM with an online softmax and never materializes the full matrix. v2 improves parallelism; v3 uses Hopper async and FP8."},"paper":{"title":"FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness","authors":["Tri Dao","Daniel Y. Fu","Stefano Ermon","et al."],"firstPublished":"2022-05-27","year":2022,"arxiv":"2205.14135","url":"https://arxiv.org/abs/2205.14135","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["transformer"],"usedBy":["nsa","ring-attention"],"related":["paged-attention","ring-attention"],"usedIn":{"zh":"PyTorch SDPA、vLLM、几乎所有训练/推理框架","en":"PyTorch SDPA, vLLM, most training/inference stacks"},"resources":[{"title":"Dao-AILab/flash-attention（官方实现）","url":"https://github.com/Dao-AILab/flash-attention","type":"repo","lang":"en"},{"title":"PyTorch 文档：scaled_dot_product_attention","url":"https://docs.pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html","type":"docs","lang":"en"},{"title":"Hugging Face：The Ultra-Scale Playbook","url":"https://huggingface.co/spaces/nanotron/ultrascale-playbook","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/flashattention/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/flashattention/"}},{"id":"sliding-window","name":{"zh":"滑动窗口/稀疏注意力","en":"Sliding-Window & Sparse Attention"},"aka":[],"category":"transformer","tier":"core","year":2020,"difficulty":3,"summary":{"zh":"每个 token 只看附近窗口加少量全局 token，让注意力线性增长。","en":"Each token attends to a local window plus a few global tokens for linear cost."},"explain":{"zh":"完整注意力的计算量随长度平方增长。Longformer 让大多数 token 只看邻近窗口，少数全局 token 看全文，从而处理长文档。Mistral 7B、Gemma 等用滑动窗口层与全局层交替。","en":"Full attention is quadratic. Longformer lets most tokens see a local window and a few global tokens see everything. Mistral 7B and Gemma interleave sliding-window and global layers."},"paper":{"title":"Longformer: The Long-Document Transformer","authors":["Iz Beltagy","Matthew E. Peters","Arman Cohan"],"firstPublished":"2020-04-10","year":2020,"arxiv":"2004.05150","url":"https://arxiv.org/abs/2004.05150","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["transformer"],"usedBy":["nsa","streaming-llm"],"related":["nsa","streaming-llm"],"usedIn":{"zh":"Mistral 7B、Gemma 2/3","en":"Mistral 7B, Gemma 2/3"},"resources":[{"title":"Lilian Weng：The Transformer Family v2","url":"https://lilianweng.github.io/posts/2023-01-27-the-transformer-family-v2/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/sliding-window/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/sliding-window/"}},{"id":"nsa","name":{"zh":"原生稀疏注意力 NSA","en":"Native Sparse Attention (NSA)"},"aka":[],"category":"transformer","tier":"core","year":2025,"difficulty":5,"summary":{"zh":"可训练、对硬件友好的动态稀疏注意力。","en":"Trainable, hardware-aligned dynamic sparse attention."},"explain":{"zh":"把长上下文分成三路：压缩的粗粒度块、精选的重要块和局部滑动窗口，用门控合并。从预训练起就稀疏，长文本推理和训练都显著加速。","en":"Combines three branches — compressed coarse blocks, selected top blocks and a sliding window — merged by gates. Sparse from pretraining, it speeds both long-context training and inference."},"paper":{"title":"Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention","authors":["Jingyang Yuan","Huazuo Gao","Damai Dai","et al."],"firstPublished":"2025-02-16","year":2025,"arxiv":"2502.11089","url":"https://arxiv.org/abs/2502.11089","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["sliding-window","flashattention"],"usedBy":[],"related":["mla"],"usedIn":{"zh":"DeepSeek 长上下文研究；同一团队后来在 DeepSeek-V3.2-Exp 中上线了另一种细粒度稀疏注意力 DSA","en":"DeepSeek long-context research; the same team later shipped a different fine-grained sparse attention (DSA) in DeepSeek-V3.2-Exp"},"resources":[{"title":"deepseek-ai/DeepSeek-V3","url":"https://github.com/deepseek-ai/DeepSeek-V3","type":"repo","lang":"en"},{"title":"deepseek-ai/DeepSeek-V3.2-Exp（DeepSeek Sparse Attention 官方说明）","url":"https://github.com/deepseek-ai/DeepSeek-V3.2-Exp","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/nsa/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/nsa/"}},{"id":"yarn","name":{"zh":"YaRN 上下文扩展","en":"YaRN (Context Window Extension)"},"aka":[],"category":"transformer","tier":"core","year":2023,"difficulty":4,"summary":{"zh":"通过调整 RoPE 频率，用少量训练把上下文窗口扩大数倍。","en":"Rescales RoPE frequencies to extend context with little fine-tuning."},"explain":{"zh":"直接让模型读超过训练长度的文本会崩。YaRN 对不同频率的旋转维度做不同程度的插值，并调整注意力温度，只需少量长文本微调即可把 4K 扩到 64K–128K。","en":"Models break beyond trained length. YaRN interpolates RoPE dimensions differently by frequency and rescales attention temperature, extending 4K to 64K–128K with brief fine-tuning."},"paper":{"title":"YaRN: Efficient Context Window Extension of Large Language Models","authors":["Bowen Peng","Jeffrey Quesnelle","Honglu Fan","et al."],"firstPublished":"2023-08-31","year":2023,"arxiv":"2309.00071","url":"https://arxiv.org/abs/2309.00071","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["rope"],"usedBy":[],"related":["ring-attention"],"usedIn":{"zh":"Qwen 系列长上下文、DeepSeek 长上下文扩展","en":"Long-context Qwen and DeepSeek models"},"resources":[{"title":"jquesnelle/yarn（官方实现）","url":"https://github.com/jquesnelle/yarn","type":"repo","lang":"en"},{"title":"Hugging Face 博客：You could have designed state of the art positional encoding","url":"https://huggingface.co/blog/designing-positional-encoding","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/yarn/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/yarn/"}},{"id":"ring-attention","name":{"zh":"环形注意力 Ring Attention","en":"Ring Attention"},"aka":[],"category":"transformer","tier":"core","year":2023,"difficulty":5,"summary":{"zh":"把超长序列切到多张卡上，按环传递 K/V 计算注意力。","en":"Shards very long sequences across devices, passing K/V around a ring."},"explain":{"zh":"每张卡持有序列的一段，计算与本地 K/V 的注意力，同时把 K/V 传给下一张卡，通信与计算重叠。上下文长度可随设备数线性增长，是“上下文并行”的代表。","en":"Each device holds a chunk and computes attention with local K/V while passing K/V to the next device, overlapping communication. Context scales with device count — a form of context parallelism."},"paper":{"title":"Ring Attention with Blockwise Transformers for Near-Infinite Context","authors":["Hao Liu","Matei Zaharia","Pieter Abbeel"],"firstPublished":"2023-10-03","year":2023,"arxiv":"2310.01889","url":"https://arxiv.org/abs/2310.01889","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["flashattention"],"usedBy":[],"related":["yarn","tensor-parallel"],"usedIn":{"zh":"百万 token 级上下文训练","en":"Million-token context training"},"resources":[{"title":"haoliuhl/ringattention（官方实现）","url":"https://github.com/haoliuhl/ringattention","type":"repo","lang":"en"},{"title":"Hugging Face：The Ultra-Scale Playbook","url":"https://huggingface.co/spaces/nanotron/ultrascale-playbook","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/ring-attention/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/ring-attention/"}},{"id":"linear-attention","name":{"zh":"线性注意力","en":"Linear Attention"},"aka":["Gated Linear Attention","GLA"],"category":"transformer","tier":"core","year":2020,"difficulty":4,"summary":{"zh":"用核函数近似 softmax，把注意力复杂度从平方降到线性。","en":"Approximates softmax with kernels to make attention linear in length."},"explain":{"zh":"把 softmax(QKᵀ)V 改写成 φ(Q)(φ(K)ᵀV)，先算右边就只需线性时间，并且可以写成 RNN 一样的递推形式。它是 RetNet、Gated DeltaNet 等新架构的起点。","en":"Rewrite softmax(QKᵀ)V as φ(Q)(φ(K)ᵀV); computing the right side first is linear and admits an RNN-style recurrence. The starting point for RetNet and Gated DeltaNet."},"paper":{"title":"Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention","authors":["Angelos Katharopoulos","Apoorv Vyas","Nikolaos Pappas","et al."],"firstPublished":"2020-06-29","year":2020,"arxiv":"2006.16236","url":"https://arxiv.org/abs/2006.16236","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["transformer"],"usedBy":["mamba2","rwkv","retnet","gated-deltanet"],"related":["retnet","gated-deltanet","mamba"],"usedIn":{"zh":"新架构与混合模型；i14 的 Laplace 3B 每层都用门控线性注意力（GLA），每条序列的递推状态固定 29.5 MB、与上下文长度无关（AWS 博客 2026-10-20）；2026-10 的 KronLSTM（arXiv 2610.26834）提出克罗内克 Delta 更新","en":"New architectures and hybrids; i14’s Laplace 3B uses gated linear attention in every layer with a fixed 29.5 MB recurrent state per sequence (AWS blog, 20 Oct 2026); KronLSTM (arXiv 2610.26834, Oct 2026) adds a Kronecker delta update"},"resources":[{"title":"fla-org/flash-linear-attention","url":"https://github.com/fla-org/flash-linear-attention","type":"repo","lang":"en"},{"title":"Lilian Weng：The Transformer Family v2","url":"https://lilianweng.github.io/posts/2023-01-27-the-transformer-family-v2/","type":"blog","lang":"en"},{"title":"AWS 机器学习博客：在 Trainium2 上训练与服务 i14 Laplace 3B 线性注意力模型（2026-10-20）","url":"https://aws.amazon.com/blogs/machine-learning/beyond-the-kv-cache-training-and-serving-i14s-laplace-3b-linear-attention-model-on-aws-trainium2-with-native-pytorch/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/linear-attention/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/linear-attention/"}},{"id":"sparse-moe","name":{"zh":"稀疏门控混合专家","en":"Sparsely-Gated Mixture of Experts"},"aka":[],"category":"moe","tier":"core","year":2017,"difficulty":3,"summary":{"zh":"一个门控网络为每个输入只挑少数几个“专家”子网络。","en":"A gating network routes each input to only a few expert sub-networks."},"explain":{"zh":"把一层拆成很多并列的专家，路由器给每个 token 打分只选 top-k 个专家计算。总参数可以极大，但每个 token 的计算量只相当于小模型。","en":"Split a layer into many experts; a router picks the top-k for each token. Total parameters can be huge while per-token compute stays small."},"paper":{"title":"Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer","authors":["Noam Shazeer","Azalia Mirhoseini","Krzysztof Maziarz","et al."],"firstPublished":"2017-01-23","year":2017,"arxiv":"1701.06538","url":"https://arxiv.org/abs/1701.06538","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["transformer"],"usedBy":["gshard","switch-transformer","mixtral","jamba"],"related":["switch-transformer","deepseek-moe"],"usedIn":{"zh":"现代 MoE 大模型的源头","en":"Origin of modern MoE LLMs"},"resources":[{"title":"Hugging Face 博客：Mixture of Experts Explained","url":"https://huggingface.co/blog/moe","type":"blog","lang":"en"},{"title":"Lilian Weng：How to Train Really Large Models on Many GPUs?","url":"https://lilianweng.github.io/posts/2021-09-25-train-large/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/sparse-moe/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/sparse-moe/"}},{"id":"gshard","name":{"zh":"GShard 专家并行","en":"GShard (Expert Parallelism)"},"aka":[],"category":"moe","tier":"extended","year":2020,"difficulty":4,"summary":{"zh":"把 MoE 的专家分布到上千个加速器上训练。","en":"Shards MoE experts across thousands of accelerators."},"explain":{"zh":"每台设备放一部分专家，token 通过 all-to-all 通信被送到对应专家所在设备，再送回。它确立了 MoE 的“专家并行”训练方式和容量因子等工程做法。","en":"Each device hosts some experts; tokens are sent to them via all-to-all and returned. It established expert parallelism and capacity factors."},"paper":{"title":"GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding","authors":["Dmitry Lepikhin","HyoukJoong Lee","Yuanzhong Xu","et al."],"firstPublished":"2020-06-30","year":2020,"arxiv":"2006.16668","url":"https://arxiv.org/abs/2006.16668","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["sparse-moe"],"usedBy":[],"related":["switch-transformer","zero"],"usedIn":{"zh":"大规模 MoE 训练框架","en":"Large-scale MoE training frameworks"},"resources":[{"title":"Hugging Face 博客：Mixture of Experts Explained","url":"https://huggingface.co/blog/moe","type":"blog","lang":"en"},{"title":"NVIDIA/Megatron-LM","url":"https://github.com/NVIDIA/Megatron-LM","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/gshard/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/gshard/"}},{"id":"switch-transformer","name":{"zh":"Switch Transformer","en":"Switch Transformer"},"aka":[],"category":"moe","tier":"core","year":2021,"difficulty":3,"summary":{"zh":"每个 token 只路由到 1 个专家的极简 MoE。","en":"A simplified MoE that routes each token to a single expert."},"explain":{"zh":"把 top-k 简化为 top-1，配合负载均衡损失和低精度技巧，稳定训练出万亿参数模型，同等计算下训练速度明显提升。","en":"Simplifies routing to top-1 and adds load-balancing loss and precision tricks, stably training trillion-parameter models faster at equal compute."},"paper":{"title":"Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity","authors":["William Fedus","Barret Zoph","Noam Shazeer"],"firstPublished":"2021-01-11","year":2021,"arxiv":"2101.03961","url":"https://arxiv.org/abs/2101.03961","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["sparse-moe"],"usedBy":[],"related":["mixtral","aux-free-balancing"],"usedIn":{"zh":"Google MoE 研究","en":"Google MoE research"},"resources":[{"title":"Hugging Face 博客：Mixture of Experts Explained","url":"https://huggingface.co/blog/moe","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/switch-transformer/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/switch-transformer/"}},{"id":"mixtral","name":{"zh":"Mixtral（top-2 MoE）","en":"Mixtral (Top-2 MoE)"},"aka":[],"category":"moe","tier":"core","year":2024,"difficulty":3,"summary":{"zh":"8 个专家选 2 个的开源 MoE 模型，让 MoE 走进开源社区。","en":"An open MoE that picks 2 of 8 experts per token."},"explain":{"zh":"每层有 8 个前馈专家，每个 token 用 2 个，总参数约 47B，激活约 13B，效果接近更大的稠密模型。它推动了开源 MoE 的普及。","en":"Each layer has 8 FFN experts and uses 2 per token — ~47B total, ~13B active — matching larger dense models. It popularized open MoE."},"paper":{"title":"Mixtral of Experts","authors":["Albert Q. Jiang","Alexandre Sablayrolles","Antoine Roux","et al."],"firstPublished":"2024-01-08","year":2024,"arxiv":"2401.04088","url":"https://arxiv.org/abs/2401.04088","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["sparse-moe"],"usedBy":["deepseek-moe"],"related":["deepseek-moe"],"usedIn":{"zh":"Mixtral 8x7B / 8x22B","en":"Mixtral 8x7B / 8x22B"},"resources":[{"title":"Mistral AI：Mixtral of experts","url":"https://mistral.ai/news/mixtral-of-experts/","type":"blog","lang":"en"},{"title":"Hugging Face 博客：Mixture of Experts Explained","url":"https://huggingface.co/blog/moe","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/mixtral/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/mixtral/"}},{"id":"deepseek-moe","name":{"zh":"DeepSeekMoE 细粒度专家","en":"DeepSeekMoE (Fine-Grained & Shared Experts)"},"aka":[],"category":"moe","tier":"core","year":2024,"difficulty":4,"summary":{"zh":"把专家切得更细并加入共享专家，让每个专家更专精。","en":"Finer-grained experts plus always-on shared experts for specialization."},"explain":{"zh":"把一个大专家拆成多个小专家、激活更多个，组合更灵活；另设几个所有 token 都用的“共享专家”承载通用知识，减少冗余。","en":"Split experts into many smaller ones and activate more of them for flexible combinations, plus shared experts every token uses for common knowledge."},"paper":{"title":"DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models","authors":["Damai Dai","Chengqi Deng","Chenggang Zhao","et al."],"firstPublished":"2024-01-11","year":2024,"arxiv":"2401.06066","url":"https://arxiv.org/abs/2401.06066","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["mixtral"],"usedBy":["aux-free-balancing"],"related":["aux-free-balancing","mla"],"usedIn":{"zh":"DeepSeek-V2/V3/R1、Qwen MoE 系列","en":"DeepSeek-V2/V3/R1, Qwen MoE models"},"resources":[{"title":"deepseek-ai/DeepSeek-V3","url":"https://github.com/deepseek-ai/DeepSeek-V3","type":"repo","lang":"en"},{"title":"Hugging Face 博客：Mixture of Experts Explained","url":"https://huggingface.co/blog/moe","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/deepseek-moe/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/deepseek-moe/"}},{"id":"aux-free-balancing","name":{"zh":"无辅助损失负载均衡","en":"Auxiliary-Loss-Free Load Balancing"},"aka":[],"category":"moe","tier":"core","year":2024,"difficulty":4,"summary":{"zh":"用动态偏置而非额外损失来平衡专家负载。","en":"Balances expert load with dynamic biases instead of an extra loss."},"explain":{"zh":"传统做法加负载均衡损失，会干扰主任务。该方法给每个专家的路由分数加一个可调偏置：专家过忙就调低、过闲就调高，既均衡又不伤效果。","en":"Balancing losses interfere with the main objective. Instead, add a per-expert bias to routing scores, lowered when overloaded and raised when idle."},"paper":{"title":"Auxiliary-Loss-Free Load Balancing Strategy for Mixture-of-Experts","authors":["Lean Wang","Huazuo Gao","Chenggang Zhao","et al."],"firstPublished":"2024-08-28","year":2024,"arxiv":"2408.15664","url":"https://arxiv.org/abs/2408.15664","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["deepseek-moe"],"usedBy":[],"related":["switch-transformer"],"usedIn":{"zh":"DeepSeek-V3","en":"DeepSeek-V3"},"resources":[{"title":"deepseek-ai/DeepSeek-V3","url":"https://github.com/deepseek-ai/DeepSeek-V3","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/aux-free-balancing/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/aux-free-balancing/"}},{"id":"s4","name":{"zh":"结构化状态空间 S4","en":"Structured State Spaces (S4)"},"aka":[],"category":"new-arch","tier":"extended","year":2021,"difficulty":5,"summary":{"zh":"用连续时间状态空间方程高效建模超长序列。","en":"Models very long sequences efficiently with structured state-space equations."},"explain":{"zh":"把序列看成一个线性动力系统的输入，隐藏状态按固定规则演化；特殊的矩阵结构让它既能像 RNN 一样递推，又能像卷积一样并行训练。","en":"Treat the sequence as input to a linear dynamical system; a special matrix structure lets it run as an RNN at inference and as a convolution in training."},"paper":{"title":"Efficiently Modeling Long Sequences with Structured State Spaces","authors":["Albert Gu","Karan Goel","Christopher Ré"],"firstPublished":"2021-10-31","year":2021,"arxiv":"2111.00396","url":"https://arxiv.org/abs/2111.00396","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["lstm"],"usedBy":["mamba","hyena"],"related":["mamba","hyena"],"usedIn":{"zh":"长序列建模、音频、基因组","en":"Long sequences, audio, genomics"},"resources":[{"title":"The Annotated S4","url":"https://srush.github.io/annotated-s4/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/s4/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/s4/"}},{"id":"mamba","name":{"zh":"Mamba 选择性状态空间","en":"Mamba (Selective State Spaces)"},"aka":[],"category":"new-arch","tier":"core","year":2023,"difficulty":4,"summary":{"zh":"让状态空间参数随输入变化，获得线性时间且接近 Transformer 的效果。","en":"Input-dependent state spaces with linear-time cost and Transformer-like quality."},"explain":{"zh":"S4 的参数固定，无法“选择性”记忆。Mamba 让参数随当前 token 变化，决定记什么、忘什么，并用硬件感知的并行扫描实现。推理时每步计算和显存恒定，不需要 KV Cache。","en":"S4's fixed parameters cannot selectively remember. Mamba makes them depend on the token and uses a hardware-aware parallel scan. Inference uses constant memory per step with no KV cache."},"paper":{"title":"Mamba: Linear-Time Sequence Modeling with Selective State Spaces","authors":["Albert Gu","Tri Dao"],"firstPublished":"2023-12-01","year":2023,"arxiv":"2312.00752","url":"https://arxiv.org/abs/2312.00752","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["s4"],"usedBy":["mamba2","jamba"],"related":["mamba2","jamba","gated-deltanet"],"usedIn":{"zh":"Jamba、Falcon Mamba、各类混合架构","en":"Jamba, Falcon Mamba, hybrid models"},"resources":[{"title":"state-spaces/mamba（官方实现）","url":"https://github.com/state-spaces/mamba","type":"repo","lang":"en"},{"title":"Hugging Face 博客：Get on the SSM train","url":"https://huggingface.co/blog/lbourdois/get-on-the-ssm-train","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/mamba/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/mamba/"}},{"id":"mamba2","name":{"zh":"Mamba-2 / 状态空间对偶","en":"Mamba-2 (State Space Duality)"},"aka":[],"category":"new-arch","tier":"extended","year":2024,"difficulty":5,"summary":{"zh":"证明 SSM 与一类注意力等价，并据此设计更快的 Mamba。","en":"Shows SSMs and a class of attention are dual, yielding faster Mamba."},"explain":{"zh":"论文提出“结构化状态空间对偶”框架，把选择性 SSM 写成带掩码的矩阵乘法，从而能用张量核心高效计算，状态维度也能更大。","en":"Introduces state space duality: selective SSMs equal masked matrix multiplications, enabling tensor-core-friendly computation and larger states."},"paper":{"title":"Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality","authors":["Tri Dao","Albert Gu"],"firstPublished":"2024-05-31","year":2024,"arxiv":"2405.21060","url":"https://arxiv.org/abs/2405.21060","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["mamba","linear-attention"],"usedBy":["gated-deltanet"],"related":["gated-deltanet"],"usedIn":{"zh":"混合架构大模型","en":"Hybrid-architecture LLMs"},"resources":[{"title":"state-spaces/mamba（官方实现）","url":"https://github.com/state-spaces/mamba","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/mamba2/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/mamba2/"}},{"id":"rwkv","name":{"zh":"RWKV","en":"RWKV"},"aka":[],"category":"new-arch","tier":"core","year":2023,"difficulty":4,"summary":{"zh":"能像 Transformer 并行训练、像 RNN 一样推理的模型。","en":"Trains in parallel like a Transformer and infers like an RNN."},"explain":{"zh":"用带时间衰减的线性注意力式机制取代 softmax 注意力，训练可并行，推理时只需固定大小的状态。由开源社区主导，已有多代版本。","en":"Replaces softmax attention with a time-decayed linear mechanism: parallel training, fixed-size state at inference. Community-driven, with many versions."},"paper":{"title":"RWKV: Reinventing RNNs for the Transformer Era","authors":["Bo Peng","Eric Alcaide","Quentin Anthony","et al."],"firstPublished":"2023-05-22","year":2023,"arxiv":"2305.13048","url":"https://arxiv.org/abs/2305.13048","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["linear-attention","lstm"],"usedBy":[],"related":["retnet","mamba"],"usedIn":{"zh":"RWKV 系列开源模型","en":"RWKV open models"},"resources":[{"title":"BlinkDL/RWKV-LM（官方实现）","url":"https://github.com/BlinkDL/RWKV-LM","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/rwkv/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/rwkv/"}},{"id":"retnet","name":{"zh":"RetNet 保留网络","en":"Retentive Network (RetNet)"},"aka":[],"category":"new-arch","tier":"extended","year":2023,"difficulty":4,"summary":{"zh":"可并行、可递推、可分块三种模式切换的注意力替代方案。","en":"An attention alternative with parallel, recurrent and chunkwise modes."},"explain":{"zh":"用带指数衰减的“保留”机制取代 softmax 注意力，训练时并行，推理时递推、O(1) 每步开销，长序列可分块计算。","en":"A decayed retention mechanism replaces softmax attention: parallel for training, recurrent with O(1) per-step inference, chunkwise for long inputs."},"paper":{"title":"Retentive Network: A Successor to Transformer for Large Language Models","authors":["Yutao Sun","Li Dong","Shaohan Huang","et al."],"firstPublished":"2023-07-17","year":2023,"arxiv":"2307.08621","url":"https://arxiv.org/abs/2307.08621","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["linear-attention"],"usedBy":[],"related":["gated-deltanet"],"usedIn":{"zh":"线性注意力研究","en":"Linear-attention research"},"resources":[{"title":"fla-org/flash-linear-attention","url":"https://github.com/fla-org/flash-linear-attention","type":"repo","lang":"en"},{"title":"microsoft/unilm · RetNet（官方实现）","url":"https://github.com/microsoft/unilm/tree/master/retnet","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/retnet/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/retnet/"}},{"id":"hyena","name":{"zh":"Hyena 长卷积","en":"Hyena Hierarchy"},"aka":[],"category":"new-arch","tier":"extended","year":2023,"difficulty":5,"summary":{"zh":"用隐式参数化的超长卷积和门控代替注意力。","en":"Replaces attention with implicitly parameterized long convolutions and gating."},"explain":{"zh":"用小网络生成和序列等长的卷积核，配合 FFT 快速计算，再用乘法门控混合信息，复杂度近线性。在 DNA 等超长序列上表现突出。","en":"A small network generates sequence-length filters computed via FFT, mixed with multiplicative gating, at near-linear cost. Strong on very long sequences like DNA."},"paper":{"title":"Hyena Hierarchy: Towards Larger Convolutional Language Models","authors":["Michael Poli","Stefano Massaroli","Eric Nguyen","et al."],"firstPublished":"2023-02-21","year":2023,"arxiv":"2302.10866","url":"https://arxiv.org/abs/2302.10866","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["s4"],"usedBy":[],"related":["mamba"],"usedIn":{"zh":"基因组模型（如 HyenaDNA、Evo）","en":"Genomic models (e.g., HyenaDNA, Evo)"},"resources":[{"title":"fla-org/flash-linear-attention","url":"https://github.com/fla-org/flash-linear-attention","type":"repo","lang":"en"},{"title":"HazyResearch/safari（Hyena 官方实现）","url":"https://github.com/HazyResearch/safari","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/hyena/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/hyena/"}},{"id":"xlstm","name":{"zh":"xLSTM","en":"xLSTM"},"aka":[],"category":"new-arch","tier":"extended","year":2024,"difficulty":4,"summary":{"zh":"用指数门控和矩阵记忆升级 LSTM，与 Transformer 正面竞争。","en":"Upgrades LSTM with exponential gating and matrix memory."},"explain":{"zh":"LSTM 原作者团队提出 sLSTM（指数门控）与 mLSTM（矩阵记忆、可并行）两种单元，堆叠成大模型，在语言建模上与 Transformer、Mamba 相当。","en":"LSTM's originators propose sLSTM (exponential gating) and mLSTM (parallelizable matrix memory), stacking them into LLMs competitive with Transformers and Mamba."},"paper":{"title":"xLSTM: Extended Long Short-Term Memory","authors":["Maximilian Beck","Korbinian Pöppel","Markus Spanring","et al."],"firstPublished":"2024-05-07","year":2024,"arxiv":"2405.04517","url":"https://arxiv.org/abs/2405.04517","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["lstm"],"usedBy":[],"related":["mamba","rwkv"],"usedIn":{"zh":"xLSTM 7B 等","en":"xLSTM 7B"},"resources":[{"title":"NX-AI/xlstm（官方实现）","url":"https://github.com/NX-AI/xlstm","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/xlstm/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/xlstm/"}},{"id":"gated-deltanet","name":{"zh":"门控 DeltaNet","en":"Gated Delta Networks"},"aka":[],"category":"new-arch","tier":"core","year":2024,"difficulty":5,"summary":{"zh":"把“增量规则”记忆更新和门控遗忘结合的线性注意力。","en":"Linear attention combining delta-rule memory updates with gated forgetting."},"explain":{"zh":"线性注意力把信息累加进一个记忆矩阵，容易“塞满”。Delta 规则在写入前先擦除旧值，门控负责整体遗忘，两者结合在检索和长上下文任务上优于 Mamba-2。","en":"Linear attention accumulates into a memory matrix that saturates. The delta rule overwrites old values before writing and gating handles forgetting; together they beat Mamba-2 on recall and long context."},"paper":{"title":"Gated Delta Networks: Improving Mamba2 with Delta Rule","authors":["Songlin Yang","Jan Kautz","Ali Hatamizadeh"],"firstPublished":"2024-12-09","year":2024,"arxiv":"2412.06464","url":"https://arxiv.org/abs/2412.06464","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["linear-attention","mamba2"],"usedBy":[],"related":["jamba"],"usedIn":{"zh":"Qwen3-Next 等混合线性注意力模型；2026-10 PyTorch TorchTitan 集成 Gated DeltaNet 内核用于训练；KronLSTM（2610.26834）以它为主要对比基线","en":"Hybrid linear-attention models such as Qwen3-Next; PyTorch TorchTitan integrated Gated DeltaNet kernels for training (Oct 2026); KronLSTM (2610.26834) benchmarks against it"},"resources":[{"title":"fla-org/flash-linear-attention","url":"https://github.com/fla-org/flash-linear-attention","type":"repo","lang":"en"},{"title":"NVlabs/GatedDeltaNet（官方实现）","url":"https://github.com/NVlabs/GatedDeltaNet","type":"repo","lang":"en"},{"title":"pytorch/torchtitan（PyTorch 原生大规模训练平台）","url":"https://github.com/pytorch/torchtitan","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/gated-deltanet/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/gated-deltanet/"}},{"id":"jamba","name":{"zh":"混合架构（Transformer+SSM）","en":"Hybrid Architectures (Jamba)"},"aka":[],"category":"new-arch","tier":"core","year":2024,"difficulty":4,"summary":{"zh":"在同一模型里交替堆叠注意力层和 Mamba 层（并加 MoE）。","en":"Interleaves attention and Mamba layers (plus MoE) in one model."},"explain":{"zh":"少量注意力层负责精确检索，大量 Mamba 层负责高效处理长序列，兼顾质量与吞吐。Jamba 是第一个生产级的此类模型，此后混合架构成为热门方向。","en":"A few attention layers handle precise recall while many Mamba layers process long sequences cheaply. Jamba was the first production-grade hybrid; hybrids are now a major trend."},"paper":{"title":"Jamba: A Hybrid Transformer-Mamba Language Model","authors":["Opher Lieber","Barak Lenz","Hofit Bata","et al."],"firstPublished":"2024-03-28","year":2024,"arxiv":"2403.19887","url":"https://arxiv.org/abs/2403.19887","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["mamba","sparse-moe"],"usedBy":[],"related":["gated-deltanet"],"usedIn":{"zh":"AI21 Jamba 系列","en":"AI21 Jamba models"},"resources":[{"title":"AI21 Jamba 模型卡（Hugging Face）","url":"https://huggingface.co/ai21labs/AI21-Jamba-Mini-1.5","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/jamba/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/jamba/"}},{"id":"diffusion-lm","name":{"zh":"扩散语言模型","en":"Diffusion Language Models (LLaDA)"},"aka":["Masked Diffusion Language Model","MDLM","dLLM"],"category":"new-arch","tier":"core","year":2025,"difficulty":5,"summary":{"zh":"不按从左到右生成，而是从全遮盖逐步“去噪”出整段文本。","en":"Generates text by iteratively unmasking a fully masked sequence."},"explain":{"zh":"训练时随机遮盖不同比例的 token 让模型复原；生成时从全部遮盖开始，多步并行地填出 token。LLaDA 证明了 8B 规模的扩散语言模型可以与自回归模型竞争。","en":"Train by masking random fractions of tokens; generate by starting fully masked and filling tokens in parallel over steps. LLaDA showed an 8B diffusion LM can rival autoregressive models."},"paper":{"title":"Large Language Diffusion Models","authors":["Shen Nie","Fengqi Zhu","Zebin You","et al."],"firstPublished":"2025-02-14","year":2025,"arxiv":"2502.09992","url":"https://arxiv.org/abs/2502.09992","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["bert","ddpm"],"usedBy":[],"related":["transformer"],"usedIn":{"zh":"LLaDA、MDLM 等研究；Inception 的商用扩散语言模型 Mercury 系列，2026-10-19 发布 Mercury 3（厂商自述 1M 上下文、每秒 1000+ token，未见第三方评测）；E-MoE（arXiv 2609.37533）改进少步解码","en":"LLaDA and MDLM research; Inception’s commercial Mercury dLLMs, with Mercury 3 released 19 Oct 2026 (vendor-reported 1M context, 1,000+ tokens/s, no independent tests yet); E-MoE (arXiv 2609.37533) improves few-step decoding"},"resources":[{"title":"ML-GSAI/LLaDA（官方实现）","url":"https://github.com/ML-GSAI/LLaDA","type":"repo","lang":"en"},{"title":"kuleshov-group/mdlm（MDLM 官方实现）","url":"https://github.com/kuleshov-group/mdlm","type":"repo","lang":"en"},{"title":"Inception Labs 文档：Mercury 扩散语言模型","url":"https://docs.inceptionlabs.ai/get-started/models","type":"docs","lang":"en"},{"title":"Inception 官方博客：Introducing Mercury 3（2026-10-19，厂商自述）","url":"https://www.inceptionlabs.ai/blog/introducing-mercury-3","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/diffusion-lm/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/diffusion-lm/"}},{"id":"scaling-laws","name":{"zh":"缩放定律","en":"Scaling Laws"},"aka":[],"category":"pretraining","tier":"core","year":2020,"difficulty":2,"summary":{"zh":"模型损失随参数、数据、算力按幂律平滑下降。","en":"Loss falls as a smooth power law in parameters, data and compute."},"explain":{"zh":"OpenAI 发现语言模型的损失与规模之间存在跨越多个数量级的幂律关系，因此可以用小模型实验预测大模型效果，并决定如何分配算力。这是“大力出奇迹”的理论依据。","en":"Loss follows power laws in model size, data and compute over many orders of magnitude, so small runs predict big ones and guide compute allocation."},"paper":{"title":"Scaling Laws for Neural Language Models","authors":["Jared Kaplan","Sam McCandlish","Tom Henighan","et al."],"firstPublished":"2020-01-23","year":2020,"arxiv":"2001.08361","url":"https://arxiv.org/abs/2001.08361","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["gpt3-icl"],"usedBy":["chinchilla","emergence"],"related":["chinchilla","test-time-scaling"],"usedIn":{"zh":"所有前沿模型的训练规划","en":"Training plans of every frontier lab"},"resources":[{"title":"Stanford CS336 从零构建语言模型","url":"https://cs336.stanford.edu/","type":"course","lang":"en"},{"title":"Hugging Face：Smol Training Playbook","url":"https://huggingface.co/spaces/HuggingFaceTB/smol-training-playbook","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/scaling-laws/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/scaling-laws/"}},{"id":"chinchilla","name":{"zh":"Chinchilla 计算最优","en":"Chinchilla Compute-Optimal Scaling"},"aka":[],"category":"pretraining","tier":"core","year":2022,"difficulty":3,"summary":{"zh":"同样算力下，参数量和训练 token 应同比增长（约 20 token/参数）。","en":"At fixed compute, scale parameters and tokens together (~20 tokens/param)."},"explain":{"zh":"DeepMind 发现此前大模型普遍“参数太多、数据太少”。70B 的 Chinchilla 用 1.4 万亿 token 训练，击败了 280B 的 Gopher。如今许多模型会远超这一比例以降低推理成本。","en":"DeepMind found prior LLMs over-sized and under-trained; 70B Chinchilla on 1.4T tokens beat 280B Gopher. Many models now overtrain beyond this ratio to cut inference cost."},"paper":{"title":"Training Compute-Optimal Large Language Models","authors":["Jordan Hoffmann","Sebastian Borgeaud","Arthur Mensch","et al."],"firstPublished":"2022-03-29","year":2022,"arxiv":"2203.15556","url":"https://arxiv.org/abs/2203.15556","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["scaling-laws"],"usedBy":[],"related":["mup"],"usedIn":{"zh":"大模型训练数据量规划","en":"Data budgeting for LLM training"},"resources":[{"title":"Google DeepMind：Chinchilla 博客","url":"https://deepmind.google/blog/an-empirical-analysis-of-compute-optimal-large-language-model-training/","type":"blog","lang":"en"},{"title":"Stanford CS336 从零构建语言模型","url":"https://cs336.stanford.edu/","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/chinchilla/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/chinchilla/"}},{"id":"emergence","name":{"zh":"涌现能力","en":"Emergent Abilities"},"aka":[],"category":"pretraining","tier":"core","year":2022,"difficulty":2,"summary":{"zh":"某些能力在模型变大到某个规模后才“突然”出现。","en":"Some abilities appear abruptly only past a certain scale."},"explain":{"zh":"论文统计了多步算术、多语言问答等任务，在小模型上接近随机，到一定规模后迅速提升。后续研究指出这部分源于评测指标的不连续，因此“涌现”仍有争议。","en":"Tasks like multi-step arithmetic stay near random for small models then jump at scale. Later work argued this is partly an artifact of discontinuous metrics, so emergence remains debated."},"paper":{"title":"Emergent Abilities of Large Language Models","authors":["Jason Wei","Yi Tay","Rishi Bommasani","et al."],"firstPublished":"2022-06-15","year":2022,"arxiv":"2206.07682","url":"https://arxiv.org/abs/2206.07682","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["scaling-laws"],"usedBy":[],"related":["cot"],"usedIn":{"zh":"能力评估与预测讨论","en":"Capability forecasting debates"},"resources":[{"title":"Stanford CS336 从零构建语言模型","url":"https://cs336.stanford.edu/","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/emergence/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/emergence/"}},{"id":"mup","name":{"zh":"μP 超参迁移","en":"μP (Maximal Update Parametrization)"},"aka":[],"category":"pretraining","tier":"extended","year":2022,"difficulty":5,"summary":{"zh":"调好小模型的超参数，直接迁移到大模型。","en":"Tune hyperparameters on a small model and transfer them to a large one."},"explain":{"zh":"按特定规则随宽度缩放初始化和学习率，使不同宽度的模型最优超参一致，于是只需在小模型上搜参，省下大量算力。","en":"Scale initialization and learning rates with width by specific rules so optimal hyperparameters stay fixed across sizes — tune small, train big."},"paper":{"title":"Tensor Programs V: Tuning Large Neural Networks via Zero-Shot Hyperparameter Transfer","authors":["Greg Yang","Edward J. Hu","Igor Babuschkin","et al."],"firstPublished":"2022-03-07","year":2022,"arxiv":"2203.03466","url":"https://arxiv.org/abs/2203.03466","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["adamw"],"usedBy":[],"related":["chinchilla"],"usedIn":{"zh":"大模型超参搜索（如 Cerebras-GPT）","en":"LLM hyperparameter search (e.g., Cerebras-GPT)"},"resources":[{"title":"microsoft/mup（官方实现）","url":"https://github.com/microsoft/mup","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/mup/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/mup/"}},{"id":"mtp","name":{"zh":"多 token 预测","en":"Multi-Token Prediction (MTP)"},"aka":[],"category":"pretraining","tier":"core","year":2024,"difficulty":4,"summary":{"zh":"训练时让模型一次预测后面多个 token。","en":"Trains the model to predict several future tokens at once."},"explain":{"zh":"在共享主干上加多个输出头，分别预测第 1、2、3…个后续 token。训练信号更密集，提升代码等任务能力，推理时这些头还能用于推测解码加速。","en":"Extra heads on a shared trunk predict the 1st, 2nd, 3rd… next tokens. Denser training signal helps tasks like code, and the heads can drive speculative decoding."},"paper":{"title":"Better & Faster Large Language Models via Multi-token Prediction","authors":["Fabian Gloeckle","Badr Youbi Idrissi","Baptiste Rozière","et al."],"firstPublished":"2024-04-30","year":2024,"arxiv":"2404.19737","url":"https://arxiv.org/abs/2404.19737","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["transformer"],"usedBy":[],"related":["speculative-decoding","medusa"],"usedIn":{"zh":"DeepSeek-V3","en":"DeepSeek-V3"},"resources":[{"title":"deepseek-ai/DeepSeek-V3","url":"https://github.com/deepseek-ai/DeepSeek-V3","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/mtp/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/mtp/"}},{"id":"mixed-precision","name":{"zh":"混合精度与 FP8 训练","en":"Mixed Precision & FP8 Training"},"aka":[],"category":"pretraining","tier":"core","year":2017,"difficulty":3,"summary":{"zh":"用 16 位甚至 8 位浮点计算，用 32 位保存主权重。","en":"Compute in 16- or 8-bit floats while keeping FP32 master weights."},"explain":{"zh":"低精度乘法更快、更省显存；保留 FP32 主权重并做损失缩放，避免精度损失。FP8 进一步把速度翻倍，但需要细粒度缩放策略。","en":"Low-precision math is faster and lighter; FP32 master weights and loss scaling prevent degradation. FP8 doubles throughput again but needs fine-grained scaling."},"paper":{"title":"Mixed Precision Training","authors":["Paulius Micikevicius","Sharan Narang","Jonah Alben","et al."],"firstPublished":"2017-10-10","year":2017,"arxiv":"1710.03740","url":"https://arxiv.org/abs/1710.03740","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["sgd"],"usedBy":["zero","quantization-int8"],"related":["zero","quantization-int8"],"usedIn":{"zh":"几乎所有大模型训练；DeepSeek-V3 使用 FP8","en":"Nearly all LLM training; DeepSeek-V3 uses FP8"},"resources":[{"title":"PyTorch 文档：自动混合精度","url":"https://docs.pytorch.org/docs/stable/amp.html","type":"docs","lang":"en"},{"title":"Hugging Face：The Ultra-Scale Playbook","url":"https://huggingface.co/spaces/nanotron/ultrascale-playbook","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/mixed-precision/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/mixed-precision/"}},{"id":"zero","name":{"zh":"ZeRO / FSDP 数据并行","en":"ZeRO / FSDP (Sharded Data Parallelism)"},"aka":[],"category":"pretraining","tier":"core","year":2019,"difficulty":4,"summary":{"zh":"把优化器状态、梯度和参数切片到各卡，消除数据并行的冗余。","en":"Shards optimizer states, gradients and parameters across GPUs."},"explain":{"zh":"普通数据并行每张卡都存一整份模型和优化器状态。ZeRO 分三级把它们切开，需要时再通信取回，单卡显存需求随卡数下降。PyTorch FSDP 是同类实现。","en":"Plain data parallelism replicates everything per GPU. ZeRO's three stages shard states, gradients and parameters, fetching them on demand. PyTorch FSDP is the equivalent."},"paper":{"title":"ZeRO: Memory Optimizations Toward Training Trillion Parameter Models","authors":["Samyam Rajbhandari","Jeff Rasley","Olatunji Ruwase","et al."],"firstPublished":"2019-10-04","year":2019,"arxiv":"1910.02054","url":"https://arxiv.org/abs/1910.02054","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["mixed-precision"],"usedBy":[],"related":["tensor-parallel","pipeline-parallel"],"usedIn":{"zh":"DeepSpeed、PyTorch FSDP","en":"DeepSpeed, PyTorch FSDP"},"resources":[{"title":"DeepSpeed 官方网站","url":"https://www.deepspeed.ai/","type":"docs","lang":"en"},{"title":"PyTorch 教程：FSDP","url":"https://docs.pytorch.org/tutorials/intermediate/FSDP_tutorial.html","type":"docs","lang":"en"},{"title":"Hugging Face：The Ultra-Scale Playbook","url":"https://huggingface.co/spaces/nanotron/ultrascale-playbook","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/zero/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/zero/"}},{"id":"tensor-parallel","name":{"zh":"张量并行（Megatron）","en":"Tensor Parallelism (Megatron-LM)"},"aka":[],"category":"pretraining","tier":"core","year":2019,"difficulty":4,"summary":{"zh":"把一层的矩阵乘法切到多张卡上同时算。","en":"Splits a layer's matrix multiplications across GPUs."},"explain":{"zh":"把注意力头和前馈层的权重按列/行切开，每张卡算一部分，再用少量通信合并结果。适合同一台机器内高速互联的 GPU。","en":"Split attention heads and FFN weights by columns/rows so each GPU computes a part, then combine with a few collectives. Best within a fast-interconnect node."},"paper":{"title":"Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism","authors":["Mohammad Shoeybi","Mostofa Patwary","Raul Puri","et al."],"firstPublished":"2019-09-17","year":2019,"arxiv":"1909.08053","url":"https://arxiv.org/abs/1909.08053","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["transformer"],"usedBy":["pipeline-parallel"],"related":["zero","pipeline-parallel"],"usedIn":{"zh":"Megatron-LM、NeMo 等","en":"Megatron-LM, NeMo"},"resources":[{"title":"NVIDIA/Megatron-LM","url":"https://github.com/NVIDIA/Megatron-LM","type":"repo","lang":"en"},{"title":"Hugging Face：The Ultra-Scale Playbook","url":"https://huggingface.co/spaces/nanotron/ultrascale-playbook","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/tensor-parallel/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/tensor-parallel/"}},{"id":"pipeline-parallel","name":{"zh":"流水线并行","en":"Pipeline Parallelism (GPipe)"},"aka":[],"category":"pretraining","tier":"core","year":2018,"difficulty":4,"summary":{"zh":"把模型按层切段放到不同卡上，像流水线一样处理微批次。","en":"Splits layers across devices and streams micro-batches like an assembly line."},"explain":{"zh":"每张卡负责若干层，一个 batch 被切成多个微批次依次流过，减少设备空闲（“气泡”）。与张量并行、数据并行组合成“3D 并行”。","en":"Each device owns some layers and micro-batches flow through to reduce idle bubbles. Combined with tensor and data parallelism it forms 3D parallelism."},"paper":{"title":"GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism","authors":["Yanping Huang","Youlong Cheng","Ankur Bapna","et al."],"firstPublished":"2018-11-16","year":2018,"arxiv":"1811.06965","url":"https://arxiv.org/abs/1811.06965","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["tensor-parallel"],"usedBy":[],"related":["zero"],"usedIn":{"zh":"超大模型训练","en":"Training of very large models"},"resources":[{"title":"Hugging Face：The Ultra-Scale Playbook","url":"https://huggingface.co/spaces/nanotron/ultrascale-playbook","type":"book","lang":"en"},{"title":"NVIDIA/Megatron-LM","url":"https://github.com/NVIDIA/Megatron-LM","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/pipeline-parallel/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/pipeline-parallel/"}},{"id":"rlhf","name":{"zh":"人类反馈强化学习 RLHF","en":"Reinforcement Learning from Human Feedback (RLHF)"},"aka":[],"category":"alignment","tier":"core","year":2022,"difficulty":3,"summary":{"zh":"先用人类偏好训练奖励模型，再用强化学习让模型迎合偏好。","en":"Train a reward model on human preferences, then optimize the policy with RL."},"explain":{"zh":"三步走：监督微调（SFT）→ 人类对多条回答排序、训练奖励模型 → 用 PPO 最大化奖励，同时用 KL 惩罚防止偏离太远。InstructGPT/ChatGPT 靠它变得“听话”。","en":"Three stages: supervised fine-tuning, a reward model trained on human rankings, then PPO to maximize reward with a KL penalty. It made InstructGPT/ChatGPT follow instructions."},"paper":{"title":"Training language models to follow instructions with human feedback","authors":["Long Ouyang","Jeff Wu","Xu Jiang","et al."],"firstPublished":"2022-03-04","year":2022,"arxiv":"2203.02155","url":"https://arxiv.org/abs/2203.02155","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["ppo","instruction-tuning"],"usedBy":["dpo","constitutional-ai","grpo","gcg-jailbreak","red-teaming","llm-judge","weak-to-strong","debate","sleeper-agents"],"related":["dpo","constitutional-ai","grpo"],"usedIn":{"zh":"ChatGPT、Claude、Llama 2-Chat 等","en":"ChatGPT, Claude, Llama 2-Chat"},"resources":[{"title":"Hugging Face 博客：Illustrating RLHF","url":"https://huggingface.co/blog/rlhf","type":"blog","lang":"en"},{"title":"Nathan Lambert：RLHF Book","url":"https://rlhfbook.com/","type":"book","lang":"en"},{"title":"TRL：PPO Trainer","url":"https://huggingface.co/docs/trl/ppo_trainer","type":"docs","lang":"en"},{"title":"OpenRLHF","url":"https://github.com/OpenRLHF/OpenRLHF","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/rlhf/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/rlhf/"}},{"id":"reward-model","name":{"zh":"奖励模型","en":"Reward Modeling (Preference Learning)"},"aka":[],"category":"alignment","tier":"core","year":2017,"difficulty":3,"summary":{"zh":"从“A 比 B 好”的比较数据中学出一个打分函数。","en":"Learns a scoring function from pairwise human comparisons."},"explain":{"zh":"用 Bradley-Terry 模型把成对偏好转成损失：被偏好回答的分数应高于另一条。它是 RLHF 的核心，也会被“奖励黑客”钻空子。","en":"A Bradley–Terry loss pushes the preferred answer's score above the other. It is central to RLHF and vulnerable to reward hacking."},"paper":{"title":"Deep reinforcement learning from human preferences","authors":["Paul Christiano","Jan Leike","Tom B. Brown","et al."],"firstPublished":"2017-06-12","year":2017,"arxiv":"1706.03741","url":"https://arxiv.org/abs/1706.03741","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["logistic-regression"],"usedBy":["prm"],"related":["rlhf","prm"],"usedIn":{"zh":"RLHF 流水线、最佳-N 采样","en":"RLHF pipelines, best-of-N reranking"},"resources":[{"title":"Nathan Lambert：RLHF Book","url":"https://rlhfbook.com/","type":"book","lang":"en"},{"title":"Lilian Weng：Reward Hacking in Reinforcement Learning","url":"https://lilianweng.github.io/posts/2024-11-28-reward-hacking/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/reward-model/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/reward-model/"}},{"id":"dpo","name":{"zh":"直接偏好优化 DPO","en":"Direct Preference Optimization (DPO)"},"aka":[],"category":"alignment","tier":"core","year":2023,"difficulty":3,"summary":{"zh":"不训练奖励模型、不跑强化学习，直接用偏好数据微调。","en":"Optimizes preferences directly — no reward model or RL loop."},"explain":{"zh":"作者证明 RLHF 的最优策略可以用一个闭式关系表达，于是把偏好损失直接写在策略模型上：提高被偏好回答相对参考模型的概率。简单稳定，开源模型广泛使用。","en":"RLHF's optimal policy has a closed form, so the preference loss can be written on the policy itself: raise the preferred answer's likelihood relative to a reference model. Simple, stable and widely used."},"paper":{"title":"Direct Preference Optimization: Your Language Model is Secretly a Reward Model","authors":["Rafael Rafailov","Archit Sharma","Eric Mitchell","et al."],"firstPublished":"2023-05-29","year":2023,"arxiv":"2305.18290","url":"https://arxiv.org/abs/2305.18290","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["rlhf"],"usedBy":["kto","orpo","simpo"],"related":["kto","orpo","simpo"],"usedIn":{"zh":"Zephyr、Tulu、Llama 3 后训练","en":"Zephyr, Tulu, Llama 3 post-training"},"resources":[{"title":"TRL：DPO Trainer","url":"https://huggingface.co/docs/trl/dpo_trainer","type":"docs","lang":"en"},{"title":"Nathan Lambert：RLHF Book","url":"https://rlhfbook.com/","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/dpo/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/dpo/"}},{"id":"kto","name":{"zh":"KTO","en":"KTO (Kahneman–Tversky Optimization)"},"aka":[],"category":"alignment","tier":"extended","year":2024,"difficulty":3,"summary":{"zh":"只需“好/不好”单条标签，不需要成对比较的对齐方法。","en":"Aligns from simple good/bad labels instead of pairs."},"explain":{"zh":"借用前景理论的效用函数，对好样本提升、坏样本压低，且对损失更敏感。数据更易收集。","en":"Uses a prospect-theory utility to raise desirable and lower undesirable outputs, with loss aversion. Data is easier to collect."},"paper":{"title":"KTO: Model Alignment as Prospect Theoretic Optimization","authors":["Kawin Ethayarajh","Winnie Xu","Niklas Muennighoff","et al."],"firstPublished":"2024-02-02","year":2024,"arxiv":"2402.01306","url":"https://arxiv.org/abs/2402.01306","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["dpo"],"usedBy":[],"related":["orpo"],"usedIn":{"zh":"偏好数据稀缺的业务对齐","en":"Alignment when only thumbs-up/down data exists"},"resources":[{"title":"Hugging Face TRL 文档","url":"https://huggingface.co/docs/trl/index","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/kto/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/kto/"}},{"id":"orpo","name":{"zh":"ORPO","en":"ORPO (Odds Ratio Preference Optimization)"},"aka":[],"category":"alignment","tier":"extended","year":2024,"difficulty":3,"summary":{"zh":"把 SFT 和偏好优化合成一步、无需参考模型。","en":"Merges SFT and preference optimization in one step without a reference model."},"explain":{"zh":"在普通 SFT 损失上加一项比较被偏好与被拒绝回答“几率比”的惩罚，单阶段完成对齐，省显存。","en":"Adds an odds-ratio penalty between chosen and rejected answers to the SFT loss — single-stage, reference-free."},"paper":{"title":"ORPO: Monolithic Preference Optimization without Reference Model","authors":["Jiwoo Hong","Noah Lee","James Thorne"],"firstPublished":"2024-03-12","year":2024,"arxiv":"2403.07691","url":"https://arxiv.org/abs/2403.07691","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["dpo"],"usedBy":[],"related":["simpo"],"usedIn":{"zh":"小团队对齐流水线","en":"Lightweight alignment pipelines"},"resources":[{"title":"Hugging Face TRL 文档","url":"https://huggingface.co/docs/trl/index","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/orpo/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/orpo/"}},{"id":"simpo","name":{"zh":"SimPO","en":"SimPO (Simple Preference Optimization)"},"aka":[],"category":"alignment","tier":"core","year":2024,"difficulty":3,"summary":{"zh":"用长度归一化的平均对数概率做隐式奖励，无需参考模型。","en":"Uses length-normalized log-likelihood as an implicit reward, reference-free."},"explain":{"zh":"DPO 需要参考模型且偏爱长回答。SimPO 用平均对数概率当奖励并加一个目标间隔，更省、更不易“越写越长”。","en":"DPO needs a reference model and favors length; SimPO uses average log-probability plus a margin, cheaper and less verbose."},"paper":{"title":"SimPO: Simple Preference Optimization with a Reference-Free Reward","authors":["Yu Meng","Mengzhou Xia","Danqi Chen"],"firstPublished":"2024-05-23","year":2024,"arxiv":"2405.14734","url":"https://arxiv.org/abs/2405.14734","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["dpo"],"usedBy":[],"related":["orpo"],"usedIn":{"zh":"开源模型偏好微调","en":"Open-model preference tuning"},"resources":[{"title":"Hugging Face TRL 文档","url":"https://huggingface.co/docs/trl/index","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/simpo/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/simpo/"}},{"id":"constitutional-ai","name":{"zh":"宪法式 AI / RLAIF","en":"Constitutional AI / RLAIF"},"aka":[],"category":"alignment","tier":"core","year":2022,"difficulty":3,"summary":{"zh":"用一份成文“宪法”让 AI 自己批评修改回答，用 AI 反馈代替人工标注。","en":"Uses a written constitution so AI critiques itself; AI feedback replaces human labels."},"explain":{"zh":"模型先按原则自我批评、改写有害回答做监督学习，再让 AI 按原则比较回答生成偏好数据，进行强化学习。减少了对有害内容的人工标注。","en":"The model critiques and revises harmful answers per principles for SFT, then AI-generated preferences drive RL, reducing human labeling of harmful content."},"paper":{"title":"Constitutional AI: Harmlessness from AI Feedback","authors":["Yuntao Bai","Saurav Kadavath","Sandipan Kundu","et al."],"firstPublished":"2022-12-15","year":2022,"arxiv":"2212.08073","url":"https://arxiv.org/abs/2212.08073","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["rlhf"],"usedBy":[],"related":["debate","dpo"],"usedIn":{"zh":"Anthropic Claude","en":"Anthropic Claude"},"resources":[{"title":"Anthropic：Constitutional AI 研究页","url":"https://www.anthropic.com/research/constitutional-ai-harmlessness-from-ai-feedback","type":"blog","lang":"en"},{"title":"Nathan Lambert：RLHF Book","url":"https://rlhfbook.com/","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/constitutional-ai/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/constitutional-ai/"}},{"id":"grpo","name":{"zh":"组相对策略优化 GRPO","en":"Group Relative Policy Optimization (GRPO)"},"aka":[],"category":"alignment","tier":"core","year":2024,"difficulty":4,"summary":{"zh":"同一题采样一组答案，用组内相对得分代替价值网络。","en":"Samples a group of answers per prompt and uses in-group relative scores instead of a critic."},"explain":{"zh":"PPO 需要和策略一样大的价值网络。GRPO 对每个问题生成多条回答，用“比组平均好多少”作为优势值，省掉价值网络，显存减半。DeepSeek-R1 用它做大规模推理强化学习。","en":"PPO needs a critic as large as the policy. GRPO samples several answers and uses each answer's reward relative to the group mean as advantage, dropping the critic. DeepSeek-R1 used it for reasoning RL."},"paper":{"title":"DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models","authors":["Zhihong Shao","Peiyi Wang","Qihao Zhu","et al."],"firstPublished":"2024-02-05","year":2024,"arxiv":"2402.03300","url":"https://arxiv.org/abs/2402.03300","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["ppo","rlhf"],"usedBy":["rlvr","dapo","deepseek-r1"],"related":["rlvr","dapo","deepseek-r1"],"usedIn":{"zh":"DeepSeekMath、DeepSeek-R1 和大量开源推理模型；2026-10 仍是新论文的默认基线（如文生图提示改写 SCOPE 2610.22264、推荐 ReMaRk 2610.17771、异步 RL 的 KLPO 2610.08963）","en":"DeepSeekMath, DeepSeek-R1 and many open reasoners; still the default baseline in Oct 2026 papers (e.g. SCOPE 2610.22264, ReMaRk 2610.17771, KLPO 2610.08963)"},"resources":[{"title":"TRL：GRPO Trainer","url":"https://huggingface.co/docs/trl/grpo_trainer","type":"docs","lang":"en"},{"title":"verl（字节跳动 RL 训练框架）","url":"https://github.com/verl-project/verl","type":"repo","lang":"en"},{"title":"huggingface/open-r1","url":"https://github.com/huggingface/open-r1","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/grpo/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/grpo/"}},{"id":"rlvr","name":{"zh":"可验证奖励强化学习 RLVR","en":"RL with Verifiable Rewards (RLVR)"},"aka":[],"category":"alignment","tier":"core","year":2024,"difficulty":4,"summary":{"zh":"用程序可判定对错的任务（数学答案、单元测试）直接给奖励。","en":"Rewards come from programmatic checks like math answers or unit tests."},"explain":{"zh":"不用奖励模型，而是检查最终答案是否正确或代码能否通过测试，奖励无法被“讨好”。它是 o1/R1 类推理模型训练的关键配方。名称由 Tulu 3 报告提出。","en":"Instead of a learned reward model, check the final answer or run tests — rewards that cannot be flattered. It is key to o1/R1-style reasoner training; the term comes from Tulu 3."},"paper":{"title":"Tulu 3: Pushing Frontiers in Open Language Model Post-Training","authors":["Nathan Lambert","Jacob Morrison","Valentina Pyatkin","et al."],"firstPublished":"2024-11-22","year":2024,"arxiv":"2411.15124","url":"https://arxiv.org/abs/2411.15124","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["grpo","ppo"],"usedBy":["deepseek-r1"],"related":["deepseek-r1","prm"],"usedIn":{"zh":"Tulu 3、DeepSeek-R1、Qwen3 推理训练","en":"Tulu 3, DeepSeek-R1, Qwen3 reasoning"},"resources":[{"title":"allenai/open-instruct（Tulu 3）","url":"https://github.com/allenai/open-instruct","type":"repo","lang":"en"},{"title":"verl（字节跳动 RL 训练框架）","url":"https://github.com/verl-project/verl","type":"repo","lang":"en"},{"title":"Nathan Lambert：RLHF Book","url":"https://rlhfbook.com/","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/rlvr/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/rlvr/"}},{"id":"dapo","name":{"zh":"DAPO","en":"DAPO"},"aka":[],"category":"alignment","tier":"core","year":2025,"difficulty":5,"summary":{"zh":"面向长思维链强化学习的一组 GRPO 改进与开源系统。","en":"A set of GRPO fixes and an open system for long-CoT RL."},"explain":{"zh":"提出非对称裁剪（Clip-Higher）、动态采样、token 级损失和超长回答惩罚，缓解熵坍缩和训练不稳，用 Qwen2.5-32B 在 AIME 2024 上取得 50 分。","en":"Introduces Clip-Higher, dynamic sampling, token-level loss and overlong reward shaping to fix entropy collapse, reaching 50 on AIME 2024 with Qwen2.5-32B."},"paper":{"title":"DAPO: An Open-Source LLM Reinforcement Learning System at Scale","authors":["Qiying Yu","Zheng Zhang","Ruofei Zhu","et al."],"firstPublished":"2025-03-18","year":2025,"arxiv":"2503.14476","url":"https://arxiv.org/abs/2503.14476","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["grpo"],"usedBy":[],"related":["rlvr"],"usedIn":{"zh":"开源推理强化学习（verl 内置 DAPO 配方）；Clip-Higher、动态采样等技巧已被后续 GRPO 变体广泛沿用","en":"Open reasoning RL (DAPO recipe ships in verl); Clip-Higher and dynamic sampling are widely reused by later GRPO variants"},"resources":[{"title":"verl（字节跳动 RL 训练框架）","url":"https://github.com/verl-project/verl","type":"repo","lang":"en"},{"title":"BytedTsinghua-SIA/DAPO（官方代码与数据）","url":"https://github.com/BytedTsinghua-SIA/DAPO","type":"repo","lang":"en"},{"title":"DAPO 项目主页","url":"https://dapo-sia.github.io/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/dapo/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/dapo/"}},{"id":"cot","name":{"zh":"思维链提示","en":"Chain-of-Thought Prompting"},"aka":[],"category":"reasoning","tier":"core","year":2022,"difficulty":1,"summary":{"zh":"让模型先写出推理步骤再给答案。","en":"Prompting the model to write intermediate reasoning before answering."},"explain":{"zh":"在示例里给出逐步推理过程，模型就会模仿，在数学和逻辑题上大幅提升。它也是后来“推理模型”通过强化学习延长思考的起点。","en":"Show step-by-step worked examples and the model imitates them, greatly improving math and logic. It is the root of RL-trained reasoning models."},"paper":{"title":"Chain-of-Thought Prompting Elicits Reasoning in Large Language Models","authors":["Jason Wei","Xuezhi Wang","Dale Schuurmans","et al."],"firstPublished":"2022-01-28","year":2022,"arxiv":"2201.11903","url":"https://arxiv.org/abs/2201.11903","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["gpt3-icl"],"usedBy":["self-consistency","tree-of-thoughts","prm","star","deepseek-r1","react","self-refine"],"related":["self-consistency","tree-of-thoughts","deepseek-r1"],"usedIn":{"zh":"所有推理任务的提示技巧","en":"Prompting for any reasoning task"},"resources":[{"title":"Lilian Weng：Prompt Engineering","url":"https://lilianweng.github.io/posts/2023-03-15-prompt-engineering/","type":"blog","lang":"en"},{"title":"Lilian Weng：Why We Think","url":"https://lilianweng.github.io/posts/2025-05-01-thinking/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/cot/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/cot/"}},{"id":"self-consistency","name":{"zh":"自洽性投票","en":"Self-Consistency"},"aka":[],"category":"reasoning","tier":"core","year":2022,"difficulty":2,"summary":{"zh":"采样多条思维链，取出现最多的答案。","en":"Samples many reasoning paths and takes the majority answer."},"explain":{"zh":"同一题让模型用较高温度回答多次，不同推理路线殊途同归的答案更可能正确。这是最简单的测试时计算扩展。","en":"Sample many CoT answers at higher temperature; answers reached by diverse paths are likelier correct. The simplest form of test-time scaling."},"paper":{"title":"Self-Consistency Improves Chain of Thought Reasoning in Language Models","authors":["Xuezhi Wang","Jason Wei","Dale Schuurmans","et al."],"firstPublished":"2022-03-21","year":2022,"arxiv":"2203.11171","url":"https://arxiv.org/abs/2203.11171","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["cot"],"usedBy":["test-time-scaling"],"related":["test-time-scaling"],"usedIn":{"zh":"推理评测中的 maj@k / cons@64","en":"maj@k / cons@64 in reasoning evals"},"resources":[{"title":"Lilian Weng：Why We Think","url":"https://lilianweng.github.io/posts/2025-05-01-thinking/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/self-consistency/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/self-consistency/"}},{"id":"tree-of-thoughts","name":{"zh":"思维树","en":"Tree of Thoughts"},"aka":[],"category":"reasoning","tier":"core","year":2023,"difficulty":3,"summary":{"zh":"把推理展开成树，边生成边评估、可回溯搜索。","en":"Expands reasoning as a tree searched with self-evaluation and backtracking."},"explain":{"zh":"每一步生成多个候选“想法”，让模型自评哪个有希望，用广度或深度优先搜索继续，必要时回退。在需要规划的谜题上远超普通思维链。","en":"Generate several candidate thoughts per step, self-evaluate them, and search breadth- or depth-first with backtracking. Excels at planning puzzles."},"paper":{"title":"Tree of Thoughts: Deliberate Problem Solving with Large Language Models","authors":["Shunyu Yao","Dian Yu","Jeffrey Zhao","et al."],"firstPublished":"2023-05-17","year":2023,"arxiv":"2305.10601","url":"https://arxiv.org/abs/2305.10601","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["cot"],"usedBy":[],"related":["mcts-alphazero","test-time-scaling"],"usedIn":{"zh":"规划类任务、Agent 搜索","en":"Planning tasks, agent search"},"resources":[{"title":"princeton-nlp/tree-of-thought-llm（官方实现）","url":"https://github.com/princeton-nlp/tree-of-thought-llm","type":"repo","lang":"en"},{"title":"Lilian Weng：LLM Powered Autonomous Agents","url":"https://lilianweng.github.io/posts/2023-06-23-agent/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/tree-of-thoughts/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/tree-of-thoughts/"}},{"id":"prm","name":{"zh":"过程奖励模型","en":"Process Reward Models (PRM)"},"aka":[],"category":"reasoning","tier":"core","year":2023,"difficulty":4,"summary":{"zh":"给推理的每一步打分，而不只看最终答案。","en":"Scores each reasoning step rather than only the final answer."},"explain":{"zh":"OpenAI 标注了 80 万条步骤级反馈训练 PRM，在数学题的最佳-N 选择上明显优于只看结果的奖励模型。它可用于搜索、重排和强化学习。","en":"OpenAI collected 800K step-level labels; a process-supervised model beat outcome supervision for best-of-N on MATH. Used for search, reranking and RL."},"paper":{"title":"Let's Verify Step by Step","authors":["Hunter Lightman","Vineet Kosaraju","Yura Burda","et al."],"firstPublished":"2023-05-31","year":2023,"arxiv":"2305.20050","url":"https://arxiv.org/abs/2305.20050","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["reward-model","cot"],"usedBy":["test-time-scaling"],"related":["test-time-scaling","rlvr"],"usedIn":{"zh":"数学推理重排、搜索","en":"Math reasoning reranking and search"},"resources":[{"title":"openai/prm800k","url":"https://github.com/openai/prm800k","type":"repo","lang":"en"},{"title":"Lilian Weng：Why We Think","url":"https://lilianweng.github.io/posts/2025-05-01-thinking/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/prm/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/prm/"}},{"id":"star","name":{"zh":"STaR 自举推理","en":"STaR (Self-Taught Reasoner)"},"aka":[],"category":"reasoning","tier":"core","year":2022,"difficulty":3,"summary":{"zh":"让模型用自己答对的推理过程反复训练自己。","en":"Bootstraps reasoning by fine-tuning on the model's own correct rationales."},"explain":{"zh":"模型为题目生成推理，保留最终答对的那些做微调；答错的给出正确答案让它“反推”理由。循环多轮，推理能力逐步增强，是“自我改进”路线的早期代表。","en":"Generate rationales, keep the ones with correct answers for fine-tuning, and rationalize failures given the answer. Iterating improves reasoning — an early self-improvement method."},"paper":{"title":"STaR: Bootstrapping Reasoning With Reasoning","authors":["Eric Zelikman","Yuhuai Wu","Jesse Mu","et al."],"firstPublished":"2022-03-28","year":2022,"arxiv":"2203.14465","url":"https://arxiv.org/abs/2203.14465","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["cot","instruction-tuning"],"usedBy":[],"related":["rlvr","deepseek-r1"],"usedIn":{"zh":"推理数据合成","en":"Reasoning data synthesis"},"resources":[{"title":"Lilian Weng：Why We Think","url":"https://lilianweng.github.io/posts/2025-05-01-thinking/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/star/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/star/"}},{"id":"test-time-scaling","name":{"zh":"测试时计算扩展","en":"Test-Time Compute Scaling"},"aka":[],"category":"reasoning","tier":"core","year":2024,"difficulty":4,"summary":{"zh":"推理时多花算力（多想、多采样、搜索）换取更好答案。","en":"Spending more inference compute — thinking, sampling, search — for better answers."},"explain":{"zh":"研究表明，在合适的策略下，把计算花在推理阶段（如用验证器搜索或迭代修改）有时比把模型做大 14 倍更划算。o1 显示出性能随“思考时间”平滑提升。","en":"With the right strategy, spending compute at inference (verifier-guided search, revisions) can beat a 14× larger model. o1 showed performance rising smoothly with thinking time."},"paper":{"title":"Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters","authors":["Charlie Snell","Jaehoon Lee","Kelvin Xu","et al."],"firstPublished":"2024-08-06","year":2024,"arxiv":"2408.03314","url":"https://arxiv.org/abs/2408.03314","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["self-consistency","prm"],"usedBy":["budget-forcing"],"related":["deepseek-r1","budget-forcing"],"usedIn":{"zh":"OpenAI o 系列、DeepSeek-R1、Claude 扩展思考","en":"OpenAI o-series, DeepSeek-R1, Claude extended thinking"},"resources":[{"title":"OpenAI：Learning to reason with LLMs","url":"https://openai.com/index/learning-to-reason-with-llms/","type":"blog","lang":"en"},{"title":"Lilian Weng：Why We Think","url":"https://lilianweng.github.io/posts/2025-05-01-thinking/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/test-time-scaling/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/test-time-scaling/"}},{"id":"deepseek-r1","name":{"zh":"DeepSeek-R1 推理强化学习配方","en":"DeepSeek-R1 Reasoning RL Recipe"},"aka":[],"category":"reasoning","tier":"core","year":2025,"difficulty":4,"summary":{"zh":"只用结果奖励的大规模强化学习，就能让模型学会长链推理与自我反思。","en":"Large-scale outcome-reward RL alone teaches long reasoning and self-reflection."},"explain":{"zh":"R1-Zero 直接在基座模型上用 GRPO + 规则奖励训练，自发出现更长思考和“回头检查”；R1 再加冷启动数据和多阶段训练提升可读性，并把推理能力蒸馏给小模型。","en":"R1-Zero applies GRPO with rule-based rewards directly to a base model, and longer thinking and self-checking emerge; R1 adds cold-start data and multi-stage training, then distills reasoning into small models."},"paper":{"title":"DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning","authors":[" DeepSeek-AI","Daya Guo","Dejian Yang","et al."],"firstPublished":"2025-01-22","year":2025,"arxiv":"2501.12948","url":"https://arxiv.org/abs/2501.12948","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["grpo","rlvr","cot"],"usedBy":[],"related":["test-time-scaling","knowledge-distillation"],"usedIn":{"zh":"DeepSeek-R1 及其蒸馏模型、众多开源复现","en":"DeepSeek-R1, its distilled models and many open replications"},"resources":[{"title":"deepseek-ai/DeepSeek-R1","url":"https://github.com/deepseek-ai/DeepSeek-R1","type":"repo","lang":"en"},{"title":"huggingface/open-r1","url":"https://github.com/huggingface/open-r1","type":"repo","lang":"en"},{"title":"Lilian Weng：Why We Think","url":"https://lilianweng.github.io/posts/2025-05-01-thinking/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/deepseek-r1/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/deepseek-r1/"}},{"id":"budget-forcing","name":{"zh":"预算强制（s1）","en":"Budget Forcing (s1)"},"aka":[],"category":"reasoning","tier":"extended","year":2025,"difficulty":3,"summary":{"zh":"用强行延长或截断“思考”的方法控制测试时算力。","en":"Controls test-time compute by forcing thinking to extend or stop."},"explain":{"zh":"只用 1000 条精选推理数据微调，再在模型想结束思考时追加“Wait”让它继续，或到预算时强制结束。简单方法就展示了测试时扩展。","en":"Fine-tune on 1,000 curated reasoning traces, then append “Wait” when the model tries to stop, or cut it off at budget — a minimal demo of test-time scaling."},"paper":{"title":"s1: Simple test-time scaling","authors":["Niklas Muennighoff","Zitong Yang","Weijia Shi","et al."],"firstPublished":"2025-01-31","year":2025,"arxiv":"2501.19393","url":"https://arxiv.org/abs/2501.19393","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["test-time-scaling"],"usedBy":[],"related":["deepseek-r1"],"usedIn":{"zh":"推理研究基线","en":"Reasoning research baseline"},"resources":[{"title":"simplescaling/s1（官方实现）","url":"https://github.com/simplescaling/s1","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/budget-forcing/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/budget-forcing/"}},{"id":"instruction-tuning","name":{"zh":"指令微调（FLAN）","en":"Instruction Tuning (FLAN)"},"aka":[],"category":"finetuning","tier":"core","year":2021,"difficulty":2,"summary":{"zh":"用大量“指令→回答”数据微调，让模型能零样本听懂新任务。","en":"Fine-tuning on many instruction–response tasks for zero-shot generalization."},"explain":{"zh":"把 60 多个 NLP 数据集改写成自然语言指令，微调后模型在没见过的任务上零样本表现大幅提升。它是所有聊天模型 SFT 阶段的基础。","en":"Rewriting 60+ NLP datasets as instructions and fine-tuning greatly improves zero-shot performance on unseen tasks. It underlies the SFT stage of every chat model."},"paper":{"title":"Finetuned Language Models Are Zero-Shot Learners","authors":["Jason Wei","Maarten Bosma","Vincent Y. Zhao","et al."],"firstPublished":"2021-09-03","year":2021,"arxiv":"2109.01652","url":"https://arxiv.org/abs/2109.01652","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["t5","gpt3-icl"],"usedBy":["rlhf","star","self-instruct","llava"],"related":["self-instruct","rlhf","lora"],"usedIn":{"zh":"所有对话模型的 SFT","en":"SFT of every chat model"},"resources":[{"title":"Hugging Face LLM 课程","url":"https://huggingface.co/learn/llm-course/chapter1/1","type":"course","lang":"multi"},{"title":"Hugging Face TRL 文档","url":"https://huggingface.co/docs/trl/index","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/instruction-tuning/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/instruction-tuning/"}},{"id":"self-instruct","name":{"zh":"Self-Instruct 合成指令","en":"Self-Instruct"},"aka":[],"category":"finetuning","tier":"core","year":2022,"difficulty":2,"summary":{"zh":"让模型自己生成指令数据，再用来微调自己。","en":"The model generates its own instruction data for fine-tuning."},"explain":{"zh":"从少量人写种子任务出发，让模型生成新指令和回答，过滤后用于微调。斯坦福 Alpaca 用类似方法以很低成本复现了指令跟随，开启合成数据时代。","en":"Starting from a few seed tasks, the model writes new instructions and answers, which are filtered and used for tuning. Stanford Alpaca used it cheaply, starting the synthetic-data era."},"paper":{"title":"Self-Instruct: Aligning Language Models with Self-Generated Instructions","authors":["Yizhong Wang","Yeganeh Kordi","Swaroop Mishra","et al."],"firstPublished":"2022-12-20","year":2022,"arxiv":"2212.10560","url":"https://arxiv.org/abs/2212.10560","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["instruction-tuning"],"usedBy":[],"related":["knowledge-distillation","star"],"usedIn":{"zh":"Alpaca、合成数据流水线","en":"Alpaca, synthetic data pipelines"},"resources":[{"title":"Stanford Alpaca 博客","url":"https://crfm.stanford.edu/2023/03/13/alpaca.html","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/self-instruct/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/self-instruct/"}},{"id":"adapters","name":{"zh":"Adapter 适配器","en":"Adapters"},"aka":[],"category":"finetuning","tier":"core","year":2019,"difficulty":2,"summary":{"zh":"在冻结的模型层间插入小模块，只训练这些小模块。","en":"Inserts small trainable modules between frozen layers."},"explain":{"zh":"每个 Transformer 层里插入一个“降维-非线性-升维”的瓶颈小网络，只训练它，参数量约为原模型的几个百分点，就接近全量微调效果。参数高效微调（PEFT）由此兴起。","en":"Insert a down-project/nonlinearity/up-project bottleneck in each layer and train only it — a few percent of parameters for near full fine-tuning quality. It launched PEFT."},"paper":{"title":"Parameter-Efficient Transfer Learning for NLP","authors":["Neil Houlsby","Andrei Giurgiu","Stanislaw Jastrzebski","et al."],"firstPublished":"2019-02-02","year":2019,"arxiv":"1902.00751","url":"https://arxiv.org/abs/1902.00751","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["bert"],"usedBy":["prompt-tuning","lora"],"related":["lora","prompt-tuning"],"usedIn":{"zh":"多任务模型适配","en":"Multi-task model adaptation"},"resources":[{"title":"Hugging Face PEFT 文档","url":"https://huggingface.co/docs/peft/index","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/adapters/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/adapters/"}},{"id":"prompt-tuning","name":{"zh":"前缀/提示微调","en":"Prefix & Prompt Tuning"},"aka":[],"category":"finetuning","tier":"extended","year":2021,"difficulty":3,"summary":{"zh":"只学习一段加在输入前的“软提示”向量。","en":"Learns only continuous soft-prompt vectors prepended to inputs."},"explain":{"zh":"冻结整个模型，在输入或每层前面加若干可训练向量，像一段机器才懂的提示。模型越大，效果越接近全量微调。","en":"Freeze the model and train a few vectors prepended to the input or each layer — a prompt only machines read. Larger models close the gap with full fine-tuning."},"paper":{"title":"The Power of Scale for Parameter-Efficient Prompt Tuning","authors":["Brian Lester","Rami Al-Rfou","Noah Constant"],"firstPublished":"2021-04-18","year":2021,"arxiv":"2104.08691","url":"https://arxiv.org/abs/2104.08691","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["adapters"],"usedBy":[],"related":["lora"],"usedIn":{"zh":"多任务服务共享一个基座","en":"Serving many tasks on one base model"},"resources":[{"title":"Hugging Face PEFT 文档","url":"https://huggingface.co/docs/peft/index","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/prompt-tuning/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/prompt-tuning/"}},{"id":"lora","name":{"zh":"LoRA 低秩适配","en":"LoRA (Low-Rank Adaptation)"},"aka":[],"category":"finetuning","tier":"core","year":2021,"difficulty":2,"summary":{"zh":"冻结原模型，只训练两个很小的低秩矩阵来表示权重变化。","en":"Freezes the model and learns weight updates as small low-rank matrices."},"explain":{"zh":"微调时权重变化 ΔW 被写成两个瘦矩阵 B×A（秩 r 很小），参数可减少上万倍；推理时可合并回原权重、无额外延迟。它是当今最流行的微调方法。","en":"Represent the update ΔW as the product of two thin matrices of small rank r, cutting trainable parameters by up to 10,000×; merge back for zero inference overhead. The most popular fine-tuning method."},"paper":{"title":"LoRA: Low-Rank Adaptation of Large Language Models","authors":["Edward J. Hu","Yelong Shen","Phillip Wallis","et al."],"firstPublished":"2021-06-17","year":2021,"arxiv":"2106.09685","url":"https://arxiv.org/abs/2106.09685","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["pca","adapters"],"usedBy":["qlora","dora","model-merging"],"related":["qlora","dora","model-merging"],"usedIn":{"zh":"Stable Diffusion 风格 LoRA、开源模型定制","en":"Stable Diffusion style LoRAs, custom open models"},"resources":[{"title":"Hugging Face PEFT 文档","url":"https://huggingface.co/docs/peft/index","type":"docs","lang":"en"},{"title":"microsoft/LoRA（官方实现）","url":"https://github.com/microsoft/LoRA","type":"repo","lang":"en"},{"title":"unslothai/unsloth","url":"https://github.com/unslothai/unsloth","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/lora/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/lora/"}},{"id":"qlora","name":{"zh":"QLoRA","en":"QLoRA"},"aka":[],"category":"finetuning","tier":"core","year":2023,"difficulty":3,"summary":{"zh":"把基座量化到 4 位再训练 LoRA，单张消费级显卡也能微调大模型。","en":"Fine-tunes LoRA on a 4-bit quantized base, fitting big models on one GPU."},"explain":{"zh":"提出 NF4 数据类型、双重量化和分页优化器，把 65B 模型的微调显存压到单张 48GB 卡，效果接近 16 位全量微调。","en":"Introduces NF4, double quantization and paged optimizers, fitting 65B fine-tuning on one 48GB GPU with near-16-bit quality."},"paper":{"title":"QLoRA: Efficient Finetuning of Quantized LLMs","authors":["Tim Dettmers","Artidoro Pagnoni","Ari Holtzman","et al."],"firstPublished":"2023-05-23","year":2023,"arxiv":"2305.14314","url":"https://arxiv.org/abs/2305.14314","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["lora","quantization-int8"],"usedBy":[],"related":["gptq"],"usedIn":{"zh":"个人与小团队微调（Unsloth、Axolotl 等）","en":"Hobbyist and small-team fine-tuning (Unsloth, Axolotl)"},"resources":[{"title":"artidoro/qlora（官方实现）","url":"https://github.com/artidoro/qlora","type":"repo","lang":"en"},{"title":"bitsandbytes 文档","url":"https://huggingface.co/docs/bitsandbytes/main/en/index","type":"docs","lang":"en"},{"title":"unslothai/unsloth","url":"https://github.com/unslothai/unsloth","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/qlora/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/qlora/"}},{"id":"dora","name":{"zh":"DoRA","en":"DoRA (Weight-Decomposed LoRA)"},"aka":[],"category":"finetuning","tier":"extended","year":2024,"difficulty":3,"summary":{"zh":"把权重拆成“大小”和“方向”，只用 LoRA 更新方向。","en":"Splits weights into magnitude and direction, using LoRA for direction."},"explain":{"zh":"分析发现全量微调与 LoRA 的更新模式不同。DoRA 单独学习每列权重的长度，并用 LoRA 调整方向，在同样参数下更接近全量微调。","en":"Full fine-tuning and LoRA update differently; DoRA learns per-column magnitude separately and uses LoRA for direction, closing the gap at equal parameters."},"paper":{"title":"DoRA: Weight-Decomposed Low-Rank Adaptation","authors":["Shih-Yang Liu","Chien-Yi Wang","Hongxu Yin","et al."],"firstPublished":"2024-02-14","year":2024,"arxiv":"2402.09353","url":"https://arxiv.org/abs/2402.09353","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["lora"],"usedBy":[],"related":["qlora"],"usedIn":{"zh":"PEFT 库可选方案","en":"Option in the PEFT library"},"resources":[{"title":"Hugging Face PEFT 文档","url":"https://huggingface.co/docs/peft/index","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/dora/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/dora/"}},{"id":"model-merging","name":{"zh":"模型合并与任务算术","en":"Model Merging & Task Arithmetic"},"aka":[],"category":"finetuning","tier":"core","year":2022,"difficulty":3,"summary":{"zh":"把多个微调模型的权重直接加减平均，组合它们的能力。","en":"Combines fine-tuned models by adding, subtracting or averaging weights."},"explain":{"zh":"微调后权重减去基座得到“任务向量”，相加能叠加能力、相减能去除行为；Model Soups 证明多个微调结果平均可以提升精度且不增加推理成本。","en":"Fine-tuned minus base weights gives a task vector; adding combines skills, subtracting removes behaviors. Model soups show averaging fine-tunes improves accuracy at no inference cost."},"paper":{"title":"Editing Models with Task Arithmetic","authors":["Gabriel Ilharco","Marco Tulio Ribeiro","Mitchell Wortsman","et al."],"firstPublished":"2022-12-08","year":2022,"arxiv":"2212.04089","url":"https://arxiv.org/abs/2212.04089","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["lora"],"usedBy":[],"related":["knowledge-distillation"],"usedIn":{"zh":"开源社区合并模型（mergekit）","en":"Community merged models (mergekit)"},"resources":[{"title":"arcee-ai/mergekit","url":"https://github.com/arcee-ai/mergekit","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/model-merging/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/model-merging/"}},{"id":"knowledge-distillation","name":{"zh":"知识蒸馏","en":"Knowledge Distillation"},"aka":["On-Policy Distillation","在线蒸馏","GKD"],"category":"compression","tier":"core","year":2015,"difficulty":2,"summary":{"zh":"让小模型学习大模型输出的“软概率”。","en":"A small student learns from a large teacher's soft probabilities."},"explain":{"zh":"老师模型的输出不只告诉正确答案，还包含“哪些错误答案也有点像”的信息，学生模仿这些软标签，能以更小的体积接近老师。现在常见的做法还有两种：用大模型生成的数据训练小模型；以及“在线蒸馏”（on-policy distillation）——让学生自己生成回答，再由老师逐 token 打分纠正，避免学生只学老师写过、自己却走不到的路径。","en":"A teacher’s soft outputs reveal which wrong answers are plausible; a student mimicking them approaches the teacher at a fraction of the size. Two common modern forms: training on teacher-generated data, and on-policy distillation, where the student generates its own answers and the teacher scores them token by token, so the student learns from its own mistakes."},"paper":{"title":"Distilling the Knowledge in a Neural Network","authors":["Geoffrey Hinton","Oriol Vinyals","Jeff Dean"],"firstPublished":"2015-03-09","year":2015,"arxiv":"1503.02531","url":"https://arxiv.org/abs/1503.02531","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["logistic-regression"],"usedBy":["weak-to-strong"],"related":["deepseek-r1","self-instruct"],"usedIn":{"zh":"DistilBERT、R1 蒸馏小模型、端侧模型；在线蒸馏用于推理模型后训练（如 2026-10 的 CriticOPD，arXiv 2610.20572）","en":"DistilBERT, R1-distilled models, on-device models; on-policy distillation in reasoning post-training (e.g. CriticOPD, arXiv 2610.20572, Oct 2026)"},"resources":[{"title":"Dive into Deep Learning（D2L）","url":"https://d2l.ai/","type":"book","lang":"en"},{"title":"Sebastian Raschka：Ahead of AI","url":"https://magazine.sebastianraschka.com/","type":"blog","lang":"en"},{"title":"Thinking Machines：On-Policy Distillation","url":"https://thinkingmachines.ai/blog/on-policy-distillation/","type":"blog","lang":"en"},{"title":"Hugging Face TRL：GKD Trainer（在线蒸馏）","url":"https://huggingface.co/docs/trl/main/en/gkd_trainer","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/knowledge-distillation/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/knowledge-distillation/"}},{"id":"quantization-int8","name":{"zh":"LLM.int8() 与离群值","en":"LLM.int8() & Outlier Handling"},"aka":[],"category":"compression","tier":"core","year":2022,"difficulty":3,"summary":{"zh":"8 位矩阵乘法，把少数“离群”特征单独用 16 位算。","en":"8-bit matmul with outlier features kept in 16-bit."},"explain":{"zh":"大模型激活里有少量数值极大的离群维度，直接量化会严重掉点。LLM.int8() 把这些维度拆出来用 FP16，其余用 INT8，几乎无损地减半显存。","en":"LLM activations contain a few huge outlier dimensions that break naive quantization. LLM.int8() keeps them in FP16 and quantizes the rest to INT8, halving memory almost losslessly."},"paper":{"title":"LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale","authors":["Tim Dettmers","Mike Lewis","Younes Belkada","et al."],"firstPublished":"2022-08-15","year":2022,"arxiv":"2208.07339","url":"https://arxiv.org/abs/2208.07339","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["mixed-precision"],"usedBy":["qlora","gptq","smoothquant","bitnet"],"related":["gptq","smoothquant","qlora"],"usedIn":{"zh":"bitsandbytes、Hugging Face 8 位加载","en":"bitsandbytes, Hugging Face 8-bit loading"},"resources":[{"title":"bitsandbytes 文档","url":"https://huggingface.co/docs/bitsandbytes/main/en/index","type":"docs","lang":"en"},{"title":"Transformers 文档：量化总览","url":"https://huggingface.co/docs/transformers/main/en/quantization/overview","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/quantization-int8/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/quantization-int8/"}},{"id":"gptq","name":{"zh":"GPTQ 训练后量化","en":"GPTQ"},"aka":[],"category":"compression","tier":"core","year":2022,"difficulty":4,"summary":{"zh":"利用二阶信息逐列量化权重，几小时把大模型压到 3–4 位。","en":"Second-order, column-by-column post-training weight quantization to 3–4 bits."},"explain":{"zh":"量化一列权重后，用近似 Hessian 把误差补偿到尚未量化的列上，只需少量校准数据，单卡几小时即可完成 175B 模型量化。","en":"After quantizing each column, compensate the error on remaining columns using an approximate Hessian. With little calibration data it quantizes 175B models in hours on one GPU."},"paper":{"title":"GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers","authors":["Elias Frantar","Saleh Ashkboos","Torsten Hoefler","et al."],"firstPublished":"2022-10-31","year":2022,"arxiv":"2210.17323","url":"https://arxiv.org/abs/2210.17323","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["quantization-int8"],"usedBy":["awq","sparsegpt"],"related":["awq","sparsegpt"],"usedIn":{"zh":"大量 GPTQ 格式开源权重","en":"Many GPTQ-format open weights"},"resources":[{"title":"ModelCloud/GPTQModel","url":"https://github.com/ModelCloud/GPTQModel","type":"repo","lang":"en"},{"title":"Transformers 文档：量化总览","url":"https://huggingface.co/docs/transformers/main/en/quantization/overview","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/gptq/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/gptq/"}},{"id":"awq","name":{"zh":"AWQ 激活感知量化","en":"AWQ (Activation-aware Weight Quantization)"},"aka":[],"category":"compression","tier":"core","year":2023,"difficulty":4,"summary":{"zh":"根据激活大小找出最重要的 1% 权重并加以保护。","en":"Protects the ~1% most salient weights identified by activation magnitude."},"explain":{"zh":"不同权重对输出的影响不同：对应大激活通道的权重更关键。AWQ 通过对这些通道做缩放来降低量化误差，无需反向传播，适合端侧部署。","en":"Weights tied to large-activation channels matter most. AWQ rescales those channels to reduce quantization error without backprop, ideal for edge deployment."},"paper":{"title":"AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration","authors":["Ji Lin","Jiaming Tang","Haotian Tang","et al."],"firstPublished":"2023-06-01","year":2023,"arxiv":"2306.00978","url":"https://arxiv.org/abs/2306.00978","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["gptq"],"usedBy":[],"related":["smoothquant"],"usedIn":{"zh":"vLLM/TensorRT-LLM 的 AWQ 模型","en":"AWQ models in vLLM and TensorRT-LLM"},"resources":[{"title":"mit-han-lab/llm-awq（官方实现）","url":"https://github.com/mit-han-lab/llm-awq","type":"repo","lang":"en"},{"title":"Transformers 文档：量化总览","url":"https://huggingface.co/docs/transformers/main/en/quantization/overview","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/awq/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/awq/"}},{"id":"smoothquant","name":{"zh":"SmoothQuant","en":"SmoothQuant"},"aka":[],"category":"compression","tier":"core","year":2022,"difficulty":4,"summary":{"zh":"把激活里的量化难度“搬”到权重上，实现 W8A8。","en":"Migrates quantization difficulty from activations to weights for W8A8."},"explain":{"zh":"激活有离群值难量化，权重则很平滑。对每个通道做数学等价的缩放，让两者都容易量化，从而权重和激活都用 INT8 计算，加速推理。","en":"Activations have outliers, weights are smooth. An equivalent per-channel scaling makes both easy to quantize, enabling INT8 weights and activations for faster inference."},"paper":{"title":"SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models","authors":["Guangxuan Xiao","Ji Lin","Mickael Seznec","et al."],"firstPublished":"2022-11-18","year":2022,"arxiv":"2211.10438","url":"https://arxiv.org/abs/2211.10438","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["quantization-int8"],"usedBy":[],"related":["awq"],"usedIn":{"zh":"TensorRT-LLM 等推理引擎","en":"TensorRT-LLM and other engines"},"resources":[{"title":"mit-han-lab/smoothquant（官方实现）","url":"https://github.com/mit-han-lab/smoothquant","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/smoothquant/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/smoothquant/"}},{"id":"sparsegpt","name":{"zh":"SparseGPT 一次性剪枝","en":"SparseGPT"},"aka":[],"category":"compression","tier":"extended","year":2023,"difficulty":4,"summary":{"zh":"不重新训练就把大模型一半权重剪掉。","en":"One-shot pruning of half a large model's weights without retraining."},"explain":{"zh":"把剪枝看成大规模稀疏回归问题，用近似二阶信息逐列决定剪哪些、并更新剩余权重补偿误差，175B 模型可剪 50% 而困惑度几乎不变。","en":"Treats pruning as sparse regression, using approximate second-order information to prune and update weights column-wise. 175B models reach 50% sparsity with little perplexity loss."},"paper":{"title":"SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shot","authors":["Elias Frantar","Dan Alistarh"],"firstPublished":"2023-01-02","year":2023,"arxiv":"2301.00774","url":"https://arxiv.org/abs/2301.00774","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["gptq"],"usedBy":[],"related":["bitnet"],"usedIn":{"zh":"稀疏推理研究","en":"Sparse inference research"},"resources":[{"title":"IST-DASLab/sparsegpt（官方实现）","url":"https://github.com/IST-DASLab/sparsegpt","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/sparsegpt/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/sparsegpt/"}},{"id":"bitnet","name":{"zh":"BitNet b1.58 三值模型","en":"BitNet b1.58 (Ternary LLMs)"},"aka":[],"category":"compression","tier":"core","year":2024,"difficulty":4,"summary":{"zh":"每个权重只取 −1、0、+1，从训练开始就是 1.58 位。","en":"Trains LLMs from scratch with ternary weights {−1, 0, +1}."},"explain":{"zh":"权重只有三种值，矩阵乘法几乎变成加减法，显存、能耗和延迟大幅下降。论文显示在 3B 规模可与 FP16 模型相当，需从头训练。","en":"With ternary weights matrix multiplication becomes additions, slashing memory, energy and latency. Matches FP16 at 3B scale, but must be trained from scratch."},"paper":{"title":"The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits","authors":["Shuming Ma","Hongyu Wang","Lingxiao Ma","et al."],"firstPublished":"2024-02-27","year":2024,"arxiv":"2402.17764","url":"https://arxiv.org/abs/2402.17764","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["quantization-int8"],"usedBy":[],"related":["gptq"],"usedIn":{"zh":"bitnet.cpp、端侧推理研究","en":"bitnet.cpp, edge inference research"},"resources":[{"title":"microsoft/BitNet","url":"https://github.com/microsoft/BitNet","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/bitnet/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/bitnet/"}},{"id":"kv-cache","name":{"zh":"KV Cache","en":"KV Cache"},"aka":[],"category":"inference","tier":"core","year":null,"difficulty":2,"summary":{"zh":"把已生成 token 的 Key/Value 存起来，避免每步重复计算。","en":"Stores past keys and values so each decoding step doesn't recompute them."},"explain":{"zh":"自回归生成时，每个新 token 都要和之前所有 token 做注意力。缓存之前算好的 K/V 后，每步只需算新 token，速度大增，但显存随上下文长度线性增长，这也是 GQA、MLA、PagedAttention 要解决的问题。","en":"Each new token attends to all previous ones; caching their K/V means only the new token is computed each step. Memory grows linearly with context — the problem GQA, MLA and PagedAttention address."},"paper":null,"verify":{"status":"pending","checkedOn":"2026-10-22"},"prereqs":["transformer"],"usedBy":["mqa","mla","speculative-decoding","paged-attention","continuous-batching","streaming-llm","pd-disaggregation"],"related":["paged-attention","gqa","mla","streaming-llm"],"usedIn":{"zh":"所有基于注意力的大模型推理引擎；线性注意力模型（如 Laplace 3B）用固定大小状态代替 KV Cache","en":"Every attention-based LLM inference engine; linear-attention models (e.g. Laplace 3B) replace it with a fixed-size state"},"resources":[{"title":"Transformers 文档：KV Cache","url":"https://huggingface.co/docs/transformers/main/en/kv_cache","type":"docs","lang":"en"},{"title":"Lilian Weng：Large Transformer Model Inference Optimization","url":"https://lilianweng.github.io/posts/2023-01-10-inference-optimization/","type":"blog","lang":"en"},{"title":"rasbt/LLMs-from-scratch","url":"https://github.com/rasbt/LLMs-from-scratch","type":"repo","lang":"en"},{"title":"AWS 机器学习博客：在 Trainium2 上训练与服务 i14 Laplace 3B 线性注意力模型（2026-10-20）","url":"https://aws.amazon.com/blogs/machine-learning/beyond-the-kv-cache-training-and-serving-i14s-laplace-3b-linear-attention-model-on-aws-trainium2-with-native-pytorch/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/kv-cache/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/kv-cache/"}},{"id":"sampling","name":{"zh":"温度与 Top-p 采样","en":"Temperature & Nucleus (Top-p) Sampling"},"aka":[],"category":"inference","tier":"core","year":2019,"difficulty":1,"summary":{"zh":"控制生成随机性：只从累计概率达到 p 的候选里抽样。","en":"Controls randomness by sampling only from the top-p probability mass."},"explain":{"zh":"贪心解码容易重复、无聊；纯随机又会胡言乱语。Top-p 每步只保留累计概率达到 p 的最小候选集合再抽样，温度则整体调高或调低分布的尖锐程度。","en":"Greedy decoding repeats; pure sampling rambles. Top-p keeps the smallest set of tokens covering probability p, and temperature sharpens or flattens the distribution."},"paper":{"title":"The Curious Case of Neural Text Degeneration","authors":["Ari Holtzman","Jan Buys","Li Du","et al."],"firstPublished":"2019-04-22","year":2019,"arxiv":"1904.09751","url":"https://arxiv.org/abs/1904.09751","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["transformer"],"usedBy":["speculative-decoding","watermarking"],"related":["speculative-decoding"],"usedIn":{"zh":"所有大模型 API 的 temperature / top_p 参数","en":"temperature / top_p in every LLM API"},"resources":[{"title":"Hugging Face LLM 课程","url":"https://huggingface.co/learn/llm-course/chapter1/1","type":"course","lang":"multi"},{"title":"Lilian Weng：Large Transformer Model Inference Optimization","url":"https://lilianweng.github.io/posts/2023-01-10-inference-optimization/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/sampling/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/sampling/"}},{"id":"speculative-decoding","name":{"zh":"推测解码","en":"Speculative Decoding"},"aka":[],"category":"inference","tier":"core","year":2022,"difficulty":3,"summary":{"zh":"小模型先猜几个 token，大模型一次性验证，输出分布不变。","en":"A small draft model proposes tokens that the big model verifies in one pass."},"explain":{"zh":"大模型解码受显存带宽限制、算力闲置。让小模型快速草拟 k 个 token，大模型并行验证并按特殊规则接受或拒绝，结果与只用大模型完全同分布，速度可提升 2–3 倍。","en":"Decoding is memory-bound with idle compute. A draft model proposes k tokens and the target verifies them in parallel with a rejection rule that preserves the exact output distribution, giving 2–3× speedups."},"paper":{"title":"Fast Inference from Transformers via Speculative Decoding","authors":["Yaniv Leviathan","Matan Kalman","Yossi Matias"],"firstPublished":"2022-11-30","year":2022,"arxiv":"2211.17192","url":"https://arxiv.org/abs/2211.17192","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["kv-cache","sampling"],"usedBy":["medusa","eagle"],"related":["medusa","eagle","mtp"],"usedIn":{"zh":"vLLM、SGLang、TensorRT-LLM、llama.cpp；2026-10 的 DLite（arXiv 2610.23787）针对上下文不断增长的智能体任务做“无历史”草稿","en":"vLLM, SGLang, TensorRT-LLM, llama.cpp; DLite (arXiv 2610.23787, Oct 2026) drafts without history for ever-growing agent contexts"},"resources":[{"title":"Hugging Face 博客：Assisted Generation","url":"https://huggingface.co/blog/assisted-generation","type":"blog","lang":"en"},{"title":"Lilian Weng：Large Transformer Model Inference Optimization","url":"https://lilianweng.github.io/posts/2023-01-10-inference-optimization/","type":"blog","lang":"en"},{"title":"vLLM 文档","url":"https://docs.vllm.ai/en/latest/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/speculative-decoding/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/speculative-decoding/"}},{"id":"medusa","name":{"zh":"Medusa 多解码头","en":"Medusa"},"aka":[],"category":"inference","tier":"extended","year":2024,"difficulty":4,"summary":{"zh":"在模型上加几个额外的解码头，自己给自己做推测解码。","en":"Adds extra decoding heads so the model drafts its own speculative tokens."},"explain":{"zh":"不需要单独的草稿模型：额外的头同时预测后面第 2、3…个 token，用树状注意力一次验证多个候选组合。","en":"No separate draft model: extra heads predict tokens 2, 3, … ahead and tree attention verifies many candidates at once."},"paper":{"title":"Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads","authors":["Tianle Cai","Yuhong Li","Zhengyang Geng","et al."],"firstPublished":"2024-01-19","year":2024,"arxiv":"2401.10774","url":"https://arxiv.org/abs/2401.10774","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["speculative-decoding"],"usedBy":[],"related":["eagle","mtp"],"usedIn":{"zh":"推理加速框架","en":"Inference acceleration frameworks"},"resources":[{"title":"FasterDecoding/Medusa（官方实现）","url":"https://github.com/FasterDecoding/Medusa","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/medusa/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/medusa/"}},{"id":"eagle","name":{"zh":"EAGLE","en":"EAGLE"},"aka":[],"category":"inference","tier":"core","year":2024,"difficulty":4,"summary":{"zh":"在特征层面而非 token 层面做推测，接受率更高。","en":"Speculates at the feature level for higher acceptance rates."},"explain":{"zh":"用一个轻量网络根据大模型倒数第二层特征和已知下一个 token 来预测后续特征，再解码成草稿 token，比 Medusa 更准，加速比更高。","en":"A light network predicts future second-to-top-layer features from current features and the next token, then decodes drafts — more accurate than Medusa."},"paper":{"title":"EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty","authors":["Yuhui Li","Fangyun Wei","Chao Zhang","et al."],"firstPublished":"2024-01-26","year":2024,"arxiv":"2401.15077","url":"https://arxiv.org/abs/2401.15077","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["speculative-decoding"],"usedBy":[],"related":["medusa"],"usedIn":{"zh":"vLLM、SGLang 的 EAGLE 推测解码","en":"EAGLE speculative decoding in vLLM and SGLang"},"resources":[{"title":"SafeAILab/EAGLE（官方实现）","url":"https://github.com/SafeAILab/EAGLE","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/eagle/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/eagle/"}},{"id":"paged-attention","name":{"zh":"PagedAttention（vLLM）","en":"PagedAttention (vLLM)"},"aka":[],"category":"inference","tier":"core","year":2023,"difficulty":3,"summary":{"zh":"像操作系统分页一样管理 KV Cache，减少显存浪费。","en":"Manages KV cache in pages like OS virtual memory."},"explain":{"zh":"把每个请求的 KV Cache 切成固定大小的块，按需分配、可共享，不再预留整段连续显存。显存碎片几乎消失，同样硬件能同时服务更多请求，吞吐提升数倍。","en":"Split each request's KV cache into fixed-size blocks allocated on demand and shareable, instead of reserving contiguous memory. Fragmentation nearly vanishes and throughput rises several-fold."},"paper":{"title":"Efficient Memory Management for Large Language Model Serving with PagedAttention","authors":["Woosuk Kwon","Zhuohan Li","Siyuan Zhuang","et al."],"firstPublished":"2023-09-12","year":2023,"arxiv":"2309.06180","url":"https://arxiv.org/abs/2309.06180","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["kv-cache"],"usedBy":["radix-attention"],"related":["continuous-batching","radix-attention"],"usedIn":{"zh":"vLLM 及多数推理引擎","en":"vLLM and most serving engines"},"resources":[{"title":"vLLM 博客：PagedAttention 发布文","url":"https://vllm.ai/blog/2023-06-20-vllm","type":"blog","lang":"en"},{"title":"vLLM 文档","url":"https://docs.vllm.ai/en/latest/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/paged-attention/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/paged-attention/"}},{"id":"continuous-batching","name":{"zh":"连续批处理","en":"Continuous (Iteration-Level) Batching"},"aka":[],"category":"inference","tier":"core","year":2022,"difficulty":3,"summary":{"zh":"每生成一步就重新组批，完成的请求立即让位给新请求。","en":"Re-forms the batch every decoding step so finished requests leave immediately."},"explain":{"zh":"传统批处理要等最长的请求结束。按“迭代”调度后，短请求一结束就换入新请求，GPU 一直满载，延迟与吞吐同时改善。","en":"Static batches wait for the longest request. Iteration-level scheduling swaps in new requests as others finish, keeping the GPU busy."},"paper":{"title":"Orca: A Distributed Serving System for Transformer-Based Generative Models","authors":["Yu et al."],"venue":"OSDI 2022","year":2022,"url":"https://www.usenix.org/conference/osdi22/presentation/yu","source":"官方页面 HTTP/WebFetch 核实","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["kv-cache"],"usedBy":["pd-disaggregation"],"related":["paged-attention"],"usedIn":{"zh":"vLLM、TGI、SGLang","en":"vLLM, TGI, SGLang"},"resources":[{"title":"vLLM 文档","url":"https://docs.vllm.ai/en/latest/","type":"docs","lang":"en"},{"title":"Lilian Weng：Large Transformer Model Inference Optimization","url":"https://lilianweng.github.io/posts/2023-01-10-inference-optimization/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/continuous-batching/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/continuous-batching/"}},{"id":"radix-attention","name":{"zh":"RadixAttention 前缀缓存","en":"RadixAttention (Prefix Caching)"},"aka":[],"category":"inference","tier":"extended","year":2023,"difficulty":3,"summary":{"zh":"用基数树自动复用不同请求间相同前缀的 KV Cache。","en":"Reuses KV cache across requests sharing prefixes via a radix tree."},"explain":{"zh":"多轮对话、Agent、少样本提示里大量请求共享相同开头。SGLang 把 KV Cache 组织成基数树，自动命中共享前缀，省掉重复计算。","en":"Chats, agents and few-shot prompts share prefixes. SGLang stores KV cache in a radix tree to reuse shared prefixes automatically."},"paper":{"title":"SGLang: Efficient Execution of Structured Language Model Programs","authors":["Lianmin Zheng","Liangsheng Yin","Zhiqiang Xie","et al."],"firstPublished":"2023-12-12","year":2023,"arxiv":"2312.07104","url":"https://arxiv.org/abs/2312.07104","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["paged-attention"],"usedBy":[],"related":["kv-cache"],"usedIn":{"zh":"SGLang；各家 API 的提示缓存","en":"SGLang; prompt caching in commercial APIs"},"resources":[{"title":"sgl-project/sglang","url":"https://github.com/sgl-project/sglang","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/radix-attention/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/radix-attention/"}},{"id":"streaming-llm","name":{"zh":"注意力汇聚点 StreamingLLM","en":"Attention Sinks (StreamingLLM)"},"aka":[],"category":"inference","tier":"extended","year":2023,"difficulty":3,"summary":{"zh":"保留开头几个“汇聚”token 的 KV，就能无限长流式生成。","en":"Keeping a few initial sink tokens enables endless streaming generation."},"explain":{"zh":"作者发现模型会把大量注意力分配给最开头的几个 token。滑动窗口丢掉它们会崩溃；只要保留这几个 token 加最近窗口，就能稳定处理数百万 token 的流。","en":"Models dump attention on the first few tokens; dropping them in a sliding window breaks generation. Keeping them plus a recent window streams millions of tokens stably."},"paper":{"title":"Efficient Streaming Language Models with Attention Sinks","authors":["Guangxuan Xiao","Yuandong Tian","Beidi Chen","et al."],"firstPublished":"2023-09-29","year":2023,"arxiv":"2309.17453","url":"https://arxiv.org/abs/2309.17453","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["kv-cache","sliding-window"],"usedBy":[],"related":["interp-sae"],"usedIn":{"zh":"长对话流式推理","en":"Long streaming conversations"},"resources":[{"title":"mit-han-lab/streaming-llm（官方实现）","url":"https://github.com/mit-han-lab/streaming-llm","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/streaming-llm/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/streaming-llm/"}},{"id":"pd-disaggregation","name":{"zh":"预填充/解码分离","en":"Prefill–Decode Disaggregation"},"aka":[],"category":"inference","tier":"extended","year":2024,"difficulty":4,"summary":{"zh":"把“读提示”和“逐字生成”放到不同 GPU 上分别优化。","en":"Runs prompt prefill and token decoding on separate GPU pools."},"explain":{"zh":"预填充是算力密集，解码是带宽密集，混在一起互相拖累。分离部署后两者各自扩缩容，满足首 token 延迟和每 token 延迟目标的吞吐更高。","en":"Prefill is compute-bound and decode bandwidth-bound; co-locating them causes interference. Separating them lets each scale to meet latency targets with higher goodput."},"paper":{"title":"DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving","authors":["Yinmin Zhong","Shengyu Liu","Junda Chen","et al."],"firstPublished":"2024-01-18","year":2024,"arxiv":"2401.09670","url":"https://arxiv.org/abs/2401.09670","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["kv-cache","continuous-batching"],"usedBy":[],"related":["paged-attention"],"usedIn":{"zh":"大规模在线推理集群","en":"Large serving clusters"},"resources":[{"title":"LLMServe/DistServe","url":"https://github.com/LLMServe/DistServe","type":"repo","lang":"en"},{"title":"vLLM 文档","url":"https://docs.vllm.ai/en/latest/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/pd-disaggregation/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/pd-disaggregation/"}},{"id":"vae","name":{"zh":"变分自编码器 VAE","en":"Variational Autoencoder (VAE)"},"aka":[],"category":"generative","tier":"core","year":2013,"difficulty":3,"summary":{"zh":"把数据编码成概率分布，再从中采样解码生成新数据。","en":"Encodes data as a distribution in latent space and decodes samples into new data."},"explain":{"zh":"编码器输出隐变量的均值和方差，通过“重参数化技巧”采样后解码重建；损失 = 重建误差 + 与标准正态的 KL 距离。Stable Diffusion 用 VAE 把图像压进潜空间。","en":"The encoder outputs mean and variance, samples via the reparameterization trick and decodes; loss = reconstruction + KL to a standard normal. Stable Diffusion's latent space comes from a VAE."},"paper":{"title":"Auto-Encoding Variational Bayes","authors":["Diederik P Kingma","Max Welling"],"firstPublished":"2013-12-20","year":2013,"arxiv":"1312.6114","url":"https://arxiv.org/abs/1312.6114","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["em","backprop"],"usedBy":["vq-vae","ddpm","latent-diffusion","world-models","interp-sae","normalizing-flows"],"related":["vq-vae","latent-diffusion"],"usedIn":{"zh":"Stable Diffusion 潜空间编码器","en":"Latent encoders in Stable Diffusion"},"resources":[{"title":"Lilian Weng：From Autoencoder to Beta-VAE","url":"https://lilianweng.github.io/posts/2018-08-12-vae/","type":"blog","lang":"en"},{"title":"Dive into Deep Learning（D2L）","url":"https://d2l.ai/","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/vae/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/vae/"}},{"id":"gan","name":{"zh":"生成对抗网络 GAN","en":"Generative Adversarial Network (GAN)"},"aka":[],"category":"generative","tier":"extended","year":2014,"difficulty":3,"summary":{"zh":"生成器造假、判别器鉴假，两者对抗中共同进步。","en":"A generator and a discriminator improve by competing."},"explain":{"zh":"生成器从噪声生成样本试图骗过判别器，判别器努力区分真假。博弈的均衡点是生成分布等于真实分布。曾主导图像生成（StyleGAN），后被扩散模型超越，但仍用于超分和加速。","en":"The generator turns noise into samples to fool the discriminator, which learns to tell real from fake; equilibrium matches the data distribution. Dominant before diffusion; still used for super-resolution and distillation."},"paper":{"title":"Generative Adversarial Networks","authors":["Ian J. Goodfellow","Jean Pouget-Abadie","Mehdi Mirza","et al."],"firstPublished":"2014-06-10","year":2014,"arxiv":"1406.2661","url":"https://arxiv.org/abs/1406.2661","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["backprop"],"usedBy":[],"related":["ddpm","vae"],"usedIn":{"zh":"StyleGAN、图像超分","en":"StyleGAN, super-resolution"},"resources":[{"title":"Lilian Weng：From GAN to WGAN","url":"https://lilianweng.github.io/posts/2017-08-20-gan/","type":"blog","lang":"en"},{"title":"Dive into Deep Learning（D2L）","url":"https://d2l.ai/","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/gan/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/gan/"}},{"id":"vq-vae","name":{"zh":"VQ-VAE 离散表示","en":"VQ-VAE (Neural Discrete Representation)"},"aka":[],"category":"generative","tier":"core","year":2017,"difficulty":3,"summary":{"zh":"把图像/音频编码成离散的“视觉词表”token。","en":"Encodes images or audio into discrete codebook tokens."},"explain":{"zh":"编码器输出被量化到最近的码本向量，于是连续信号变成一串离散 token，可以像文字一样用 Transformer 建模。它是图像分词器和音频编解码器的基础。","en":"Encoder outputs snap to the nearest codebook vector, turning signals into discrete tokens that Transformers can model like text. Basis of image tokenizers and audio codecs."},"paper":{"title":"Neural Discrete Representation Learning","authors":["Aaron van den Oord","Oriol Vinyals","Koray Kavukcuoglu"],"firstPublished":"2017-11-02","year":2017,"arxiv":"1711.00937","url":"https://arxiv.org/abs/1711.00937","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["vae"],"usedBy":["var","genie","valle"],"related":["var","transformer"],"usedIn":{"zh":"图像/音频 tokenizer、DALL·E 1","en":"Image/audio tokenizers, DALL·E 1"},"resources":[{"title":"Lilian Weng：From Autoencoder to Beta-VAE","url":"https://lilianweng.github.io/posts/2018-08-12-vae/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/vq-vae/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/vq-vae/"}},{"id":"ddpm","name":{"zh":"去噪扩散概率模型 DDPM","en":"Denoising Diffusion Probabilistic Models (DDPM)"},"aka":[],"category":"generative","tier":"core","year":2020,"difficulty":3,"summary":{"zh":"学会一步步把噪声还原成图像。","en":"Learns to turn noise into images step by step."},"explain":{"zh":"前向过程逐步给图像加高斯噪声直至纯噪声；训练网络预测每一步加入的噪声。生成时从纯噪声出发，反复去噪得到清晰图像。它开启了扩散模型时代。","en":"The forward process adds Gaussian noise until pure noise; a network learns to predict that noise. Generation starts from noise and denoises repeatedly. It launched the diffusion era."},"paper":{"title":"Denoising Diffusion Probabilistic Models","authors":["Jonathan Ho","Ajay Jain","Pieter Abbeel"],"firstPublished":"2020-06-19","year":2020,"arxiv":"2006.11239","url":"https://arxiv.org/abs/2006.11239","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["vae"],"usedBy":["diffusion-lm","ddim","score-sde","latent-diffusion","cfg","flow-matching"],"related":["ddim","score-sde","latent-diffusion","flow-matching"],"usedIn":{"zh":"Stable Diffusion、Imagen、DALL·E 2/3、视频生成","en":"Stable Diffusion, Imagen, DALL·E 2/3, video models"},"resources":[{"title":"Lilian Weng：What are Diffusion Models?","url":"https://lilianweng.github.io/posts/2021-07-11-diffusion-models/","type":"blog","lang":"en"},{"title":"Hugging Face 扩散模型课程","url":"https://huggingface.co/learn/diffusion-course/unit0/1","type":"course","lang":"en"},{"title":"Diffusers 文档","url":"https://huggingface.co/docs/diffusers/index","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/ddpm/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/ddpm/"}},{"id":"ddim","name":{"zh":"DDIM 快速采样","en":"DDIM"},"aka":[],"category":"generative","tier":"extended","year":2020,"difficulty":3,"summary":{"zh":"把扩散采样改成确定性过程，步数从上千降到几十。","en":"Makes diffusion sampling deterministic, cutting steps from ~1000 to tens."},"explain":{"zh":"在不改训练的前提下，构造非马尔可夫的反向过程，可以跳步采样，还能把图像反演回噪声做编辑。","en":"Without retraining, a non-Markovian reverse process allows skipping steps and inverting images back to noise for editing."},"paper":{"title":"Denoising Diffusion Implicit Models","authors":["Jiaming Song","Chenlin Meng","Stefano Ermon"],"firstPublished":"2020-10-06","year":2020,"arxiv":"2010.02502","url":"https://arxiv.org/abs/2010.02502","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["ddpm"],"usedBy":["consistency-models"],"related":["consistency-models"],"usedIn":{"zh":"Stable Diffusion 采样器","en":"Stable Diffusion samplers"},"resources":[{"title":"Diffusers 文档","url":"https://huggingface.co/docs/diffusers/index","type":"docs","lang":"en"},{"title":"Lilian Weng：What are Diffusion Models?","url":"https://lilianweng.github.io/posts/2021-07-11-diffusion-models/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/ddim/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/ddim/"}},{"id":"score-sde","name":{"zh":"分数匹配与随机微分方程","en":"Score-Based Models via SDEs"},"aka":[],"category":"generative","tier":"extended","year":2020,"difficulty":5,"summary":{"zh":"用随机微分方程统一扩散模型，生成即逆向求解方程。","en":"Unifies diffusion as SDEs; generation solves the reverse-time equation."},"explain":{"zh":"学习数据分布的“分数”（对数密度梯度），把加噪看作连续时间 SDE，生成就是逆向 SDE 或对应的概率流 ODE。它连接了扩散、分数匹配与后来的流匹配。","en":"Learn the score (gradient of log density), view noising as a continuous SDE, and generate by solving the reverse SDE or probability-flow ODE. It connects diffusion, score matching and flow matching."},"paper":{"title":"Score-Based Generative Modeling through Stochastic Differential Equations","authors":["Yang Song","Jascha Sohl-Dickstein","Diederik P. Kingma","et al."],"firstPublished":"2020-11-26","year":2020,"arxiv":"2011.13456","url":"https://arxiv.org/abs/2011.13456","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["ddpm"],"usedBy":["flow-matching","consistency-models"],"related":["flow-matching"],"usedIn":{"zh":"扩散理论与高阶采样器","en":"Diffusion theory and advanced samplers"},"resources":[{"title":"Yang Song：Generative Modeling by Estimating Gradients","url":"https://yang-song.net/blog/2021/score/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/score-sde/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/score-sde/"}},{"id":"latent-diffusion","name":{"zh":"潜空间扩散（Stable Diffusion）","en":"Latent Diffusion Models"},"aka":[],"category":"generative","tier":"core","year":2021,"difficulty":3,"summary":{"zh":"在 VAE 压缩后的潜空间里做扩散，大幅降低成本。","en":"Runs diffusion in a VAE's compressed latent space to cut cost."},"explain":{"zh":"直接在像素上扩散非常昂贵。先用自编码器把图像压缩 8 倍，再在潜空间去噪，并通过交叉注意力接入文本条件。这就是开源 Stable Diffusion 的核心。","en":"Pixel diffusion is expensive. Compress images ~8× with an autoencoder, denoise in latent space and inject text via cross-attention — the core of Stable Diffusion."},"paper":{"title":"High-Resolution Image Synthesis with Latent Diffusion Models","authors":["Robin Rombach","Andreas Blattmann","Dominik Lorenz","et al."],"firstPublished":"2021-12-20","year":2021,"arxiv":"2112.10752","url":"https://arxiv.org/abs/2112.10752","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["ddpm","vae","clip"],"usedBy":["dit","controlnet"],"related":["dit","cfg","controlnet"],"usedIn":{"zh":"Stable Diffusion 系列","en":"Stable Diffusion family"},"resources":[{"title":"Jay Alammar：The Illustrated Stable Diffusion","url":"https://jalammar.github.io/illustrated-stable-diffusion/","type":"blog","lang":"en"},{"title":"CompVis/latent-diffusion（官方实现）","url":"https://github.com/CompVis/latent-diffusion","type":"repo","lang":"en"},{"title":"Diffusers 文档","url":"https://huggingface.co/docs/diffusers/index","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/latent-diffusion/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/latent-diffusion/"}},{"id":"cfg","name":{"zh":"无分类器引导 CFG","en":"Classifier-Free Guidance (CFG)"},"aka":[],"category":"generative","tier":"core","year":2022,"difficulty":3,"summary":{"zh":"同时预测有条件和无条件的结果，按比例放大差异，让图更贴合提示。","en":"Extrapolates between conditional and unconditional predictions to follow prompts."},"explain":{"zh":"训练时随机丢掉文本条件，让一个模型同时会“有条件”和“无条件”生成；采样时沿二者差值方向放大，引导强度越高越贴合提示但多样性下降。","en":"Randomly drop the condition in training so one model does both; at sampling, push along the difference. Higher guidance follows prompts more at the cost of diversity."},"paper":{"title":"Classifier-Free Diffusion Guidance","authors":["Jonathan Ho","Tim Salimans"],"firstPublished":"2022-07-26","year":2022,"arxiv":"2207.12598","url":"https://arxiv.org/abs/2207.12598","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["ddpm"],"usedBy":[],"related":["latent-diffusion"],"usedIn":{"zh":"几乎所有文生图/文生视频（guidance_scale 参数）","en":"Nearly all text-to-image/video (guidance_scale)"},"resources":[{"title":"Diffusers 文档","url":"https://huggingface.co/docs/diffusers/index","type":"docs","lang":"en"},{"title":"Lilian Weng：What are Diffusion Models?","url":"https://lilianweng.github.io/posts/2021-07-11-diffusion-models/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/cfg/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/cfg/"}},{"id":"dit","name":{"zh":"扩散 Transformer DiT","en":"Diffusion Transformer (DiT)"},"aka":[],"category":"generative","tier":"core","year":2022,"difficulty":4,"summary":{"zh":"用 Transformer 代替 U-Net 做扩散去噪骨干。","en":"Replaces the U-Net denoiser with a Transformer."},"explain":{"zh":"把潜空间图像切成 patch 当 token，用 Transformer 去噪，并用自适应 LayerNorm 注入时间步和条件。它随规模扩展得很好，Sora、SD3、FLUX 等都采用了 DiT 类架构。","en":"Patchify latents into tokens and denoise with a Transformer, injecting timestep and conditions with adaptive LayerNorm. It scales well; Sora, SD3 and FLUX use DiT-style designs."},"paper":{"title":"Scalable Diffusion Models with Transformers","authors":["William Peebles","Saining Xie"],"firstPublished":"2022-12-19","year":2022,"arxiv":"2212.09748","url":"https://arxiv.org/abs/2212.09748","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["latent-diffusion","vit"],"usedBy":[],"related":["flow-matching"],"usedIn":{"zh":"Sora、Stable Diffusion 3、FLUX、各类视频模型","en":"Sora, Stable Diffusion 3, FLUX, video models"},"resources":[{"title":"facebookresearch/DiT（官方实现）","url":"https://github.com/facebookresearch/DiT","type":"repo","lang":"en"},{"title":"OpenAI：Video generation models as world simulators","url":"https://openai.com/index/video-generation-models-as-world-simulators/","type":"blog","lang":"en"},{"title":"Lilian Weng：Diffusion Models for Video Generation","url":"https://lilianweng.github.io/posts/2024-04-12-diffusion-video/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/dit/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/dit/"}},{"id":"flow-matching","name":{"zh":"流匹配","en":"Flow Matching"},"aka":[],"category":"generative","tier":"core","year":2022,"difficulty":4,"summary":{"zh":"直接学习把噪声“流动”成数据的速度场。","en":"Learns a velocity field that flows noise into data."},"explain":{"zh":"定义从噪声到数据的简单路径（如直线），训练网络回归沿路径的速度，生成时解一个常微分方程。比扩散训练更简单、采样步数更少，已成为新一代图像和视频模型的主流目标函数。","en":"Define simple paths from noise to data (e.g., straight lines), regress the velocity along them, and sample by solving an ODE. Simpler than diffusion with fewer steps; now the mainstream objective for image and video models."},"paper":{"title":"Flow Matching for Generative Modeling","authors":["Yaron Lipman","Ricky T. Q. Chen","Heli Ben-Hamu","et al."],"firstPublished":"2022-10-06","year":2022,"arxiv":"2210.02747","url":"https://arxiv.org/abs/2210.02747","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["ddpm","score-sde"],"usedBy":["rectified-flow"],"related":["rectified-flow","dit","normalizing-flows"],"usedIn":{"zh":"Stable Diffusion 3、FLUX、Meta Movie Gen","en":"Stable Diffusion 3, FLUX, Meta Movie Gen"},"resources":[{"title":"Cambridge MLG 博客：An introduction to Flow Matching","url":"https://mlg.eng.cam.ac.uk/blog/2024/01/20/flow-matching.html","type":"blog","lang":"en"},{"title":"Flow Matching Guide and Code（Meta, arXiv:2412.06264）","url":"https://arxiv.org/abs/2412.06264","type":"paper","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/flow-matching/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/flow-matching/"}},{"id":"rectified-flow","name":{"zh":"整流（Rectified Flow）","en":"Rectified Flow"},"aka":[],"category":"generative","tier":"extended","year":2022,"difficulty":4,"summary":{"zh":"把生成路径“拉直”，少数几步甚至一步就能生成。","en":"Straightens generative paths so few or one step suffices."},"explain":{"zh":"在噪声与数据之间学习直线运输，并可通过“reflow”反复拉直轨迹，使 ODE 很少几步就能精确求解。Stable Diffusion 3 采用了整流公式。","en":"Learn straight-line transport between noise and data and iteratively reflow to straighten paths so ODEs solve in a few steps. SD3 uses the rectified-flow formulation."},"paper":{"title":"Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow","authors":["Xingchao Liu","Chengyue Gong","Qiang Liu"],"firstPublished":"2022-09-07","year":2022,"arxiv":"2209.03003","url":"https://arxiv.org/abs/2209.03003","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["flow-matching"],"usedBy":[],"related":["consistency-models"],"usedIn":{"zh":"Stable Diffusion 3、FLUX","en":"Stable Diffusion 3, FLUX"},"resources":[{"title":"Cambridge MLG 博客：An introduction to Flow Matching","url":"https://mlg.eng.cam.ac.uk/blog/2024/01/20/flow-matching.html","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/rectified-flow/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/rectified-flow/"}},{"id":"consistency-models","name":{"zh":"一致性模型","en":"Consistency Models"},"aka":[],"category":"generative","tier":"extended","year":2023,"difficulty":4,"summary":{"zh":"学会从轨迹上任一点一步跳到终点，实现一步生成。","en":"Maps any point on a trajectory directly to its endpoint for one-step generation."},"explain":{"zh":"约束模型在同一条去噪轨迹上的输出一致，于是任意噪声图都能一步映射回干净图像；可从扩散模型蒸馏或独立训练。LCM 等快速生成器基于此。","en":"Enforce consistent outputs along a denoising trajectory, so any noisy input maps to a clean image in one step; trained by distillation or from scratch. Basis of LCM-style fast generators."},"paper":{"title":"Consistency Models","authors":["Yang Song","Prafulla Dhariwal","Mark Chen","et al."],"firstPublished":"2023-03-02","year":2023,"arxiv":"2303.01469","url":"https://arxiv.org/abs/2303.01469","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["ddim","score-sde"],"usedBy":[],"related":["rectified-flow"],"usedIn":{"zh":"LCM、实时图像生成","en":"LCM, real-time image generation"},"resources":[{"title":"Diffusers 文档","url":"https://huggingface.co/docs/diffusers/index","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/consistency-models/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/consistency-models/"}},{"id":"controlnet","name":{"zh":"ControlNet","en":"ControlNet"},"aka":[],"category":"generative","tier":"core","year":2023,"difficulty":3,"summary":{"zh":"给扩散模型加上边缘、姿态、深度等精确控制条件。","en":"Adds precise spatial control (edges, pose, depth) to diffusion models."},"explain":{"zh":"复制一份 U-Net 编码器接收控制图，用“零卷积”连回原模型，训练时原模型冻结，不会被破坏。用户可用草图或骨架精确控制构图。","en":"Clone the U-Net encoder to read a control image and connect it via zero-initialized convolutions with the original frozen. Users control layout with sketches or poses."},"paper":{"title":"Adding Conditional Control to Text-to-Image Diffusion Models","authors":["Lvmin Zhang","Anyi Rao","Maneesh Agrawala"],"firstPublished":"2023-02-10","year":2023,"arxiv":"2302.05543","url":"https://arxiv.org/abs/2302.05543","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["latent-diffusion"],"usedBy":[],"related":["lora"],"usedIn":{"zh":"Stable Diffusion 生态、设计工具","en":"Stable Diffusion ecosystem, design tools"},"resources":[{"title":"lllyasviel/ControlNet（官方实现）","url":"https://github.com/lllyasviel/ControlNet","type":"repo","lang":"en"},{"title":"Diffusers 文档","url":"https://huggingface.co/docs/diffusers/index","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/controlnet/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/controlnet/"}},{"id":"var","name":{"zh":"视觉自回归 VAR","en":"Visual Autoregressive Modeling (VAR)"},"aka":[],"category":"generative","tier":"extended","year":2024,"difficulty":4,"summary":{"zh":"按“由粗到细的分辨率”而不是逐像素做自回归生成图像。","en":"Autoregressive image generation by next-scale rather than next-token prediction."},"explain":{"zh":"先生成最低分辨率的 token 图，再逐级预测更高分辨率，每级内部并行。首次让 GPT 式自回归在图像生成质量和速度上超过扩散 Transformer，获 NeurIPS 2024 最佳论文。","en":"Generate a coarse token map, then predict progressively finer scales in parallel within each. It made GPT-style autoregression beat diffusion transformers in quality and speed; NeurIPS 2024 best paper."},"paper":{"title":"Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction","authors":["Keyu Tian","Yi Jiang","Zehuan Yuan","et al."],"firstPublished":"2024-04-03","year":2024,"arxiv":"2404.02905","url":"https://arxiv.org/abs/2404.02905","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["vq-vae","transformer"],"usedBy":[],"related":["dit"],"usedIn":{"zh":"自回归图像生成研究","en":"Autoregressive image generation research"},"resources":[{"title":"FoundationVision/VAR（官方实现）","url":"https://github.com/FoundationVision/VAR","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/var/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/var/"}},{"id":"vit","name":{"zh":"视觉 Transformer ViT","en":"Vision Transformer (ViT)"},"aka":[],"category":"multimodal","tier":"core","year":2020,"difficulty":2,"summary":{"zh":"把图像切成 16×16 小块当作“单词”，直接用 Transformer 处理。","en":"Treats 16×16 image patches as words for a standard Transformer."},"explain":{"zh":"图像被切块、展平、线性投影成 token 序列，加位置编码后送入 Transformer。数据足够多时超越 CNN，如今是视觉语言模型和 DiT 的标准视觉骨干。","en":"Patches are flattened, projected into tokens with positions and fed to a Transformer. With enough data it beats CNNs and is now the standard vision backbone for VLMs and DiT."},"paper":{"title":"An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale","authors":["Alexey Dosovitskiy","Lucas Beyer","Alexander Kolesnikov","et al."],"firstPublished":"2020-10-22","year":2020,"arxiv":"2010.11929","url":"https://arxiv.org/abs/2010.11929","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["transformer","cnn"],"usedBy":["dit","clip","sam","jepa","mae","swin","dinov2"],"related":["clip","dit","mae"],"usedIn":{"zh":"CLIP、SigLIP、各类 VLM 的视觉编码器","en":"Vision encoders of CLIP, SigLIP and VLMs"},"resources":[{"title":"Dive into Deep Learning（D2L）","url":"https://d2l.ai/","type":"book","lang":"en"},{"title":"Stanford CS231n 视觉识别中的卷积网络","url":"https://cs231n.github.io/","type":"course","lang":"en"},{"title":"Hugging Face 博客：Vision Language Models Explained","url":"https://huggingface.co/blog/vlms","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/vit/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/vit/"}},{"id":"clip","name":{"zh":"CLIP 图文对比学习","en":"CLIP (Contrastive Language–Image Pretraining)"},"aka":[],"category":"multimodal","tier":"core","year":2021,"difficulty":2,"summary":{"zh":"用 4 亿图文对训练，让图像和文字落在同一个向量空间。","en":"Aligns images and text in one embedding space from 400M pairs."},"explain":{"zh":"图像编码器和文本编码器分别编码，一批里配对的图文相似度拉高、不配对的压低（对比损失）。训练后可零样本分类、以文搜图，并成为文生图模型的文本条件。","en":"Image and text encoders are trained so matched pairs in a batch score high and mismatched ones low. It enables zero-shot classification and text-image search, and conditions text-to-image models."},"paper":{"title":"Learning Transferable Visual Models From Natural Language Supervision","authors":["Alec Radford","Jong Wook Kim","Chris Hallacy","et al."],"firstPublished":"2021-02-26","year":2021,"arxiv":"2103.00020","url":"https://arxiv.org/abs/2103.00020","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["vit","contrastive-learning"],"usedBy":["latent-diffusion","siglip","llava","flamingo"],"related":["siglip","llava","latent-diffusion"],"usedIn":{"zh":"Stable Diffusion 文本编码器、以文搜图","en":"Stable Diffusion text encoder, image search"},"resources":[{"title":"openai/CLIP（官方实现）","url":"https://github.com/openai/CLIP","type":"repo","lang":"en"},{"title":"mlfoundations/open_clip","url":"https://github.com/mlfoundations/open_clip","type":"repo","lang":"en"},{"title":"Lilian Weng：Generalized Visual Language Models","url":"https://lilianweng.github.io/posts/2022-06-09-vlm/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/clip/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/clip/"}},{"id":"siglip","name":{"zh":"SigLIP","en":"SigLIP (Sigmoid Loss)"},"aka":[],"category":"multimodal","tier":"core","year":2023,"difficulty":3,"summary":{"zh":"用逐对 sigmoid 损失代替 softmax 对比损失，训练更省更稳。","en":"Replaces softmax contrastive loss with pairwise sigmoid loss."},"explain":{"zh":"每个图文对单独判断“是否匹配”，不需要在整批上做归一化，小 batch 也能训好、更易扩展。常被用作开源视觉语言模型的视觉编码器。","en":"Each image–text pair is judged independently without batch-wide normalization, working at small batches and scaling easily. A common vision encoder in open VLMs."},"paper":{"title":"Sigmoid Loss for Language Image Pre-Training","authors":["Xiaohua Zhai","Basil Mustafa","Alexander Kolesnikov","et al."],"firstPublished":"2023-03-27","year":2023,"arxiv":"2303.15343","url":"https://arxiv.org/abs/2303.15343","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["clip"],"usedBy":[],"related":["llava"],"usedIn":{"zh":"PaliGemma、Gemma 3 等 VLM","en":"PaliGemma, Gemma 3 and other VLMs"},"resources":[{"title":"mlfoundations/open_clip","url":"https://github.com/mlfoundations/open_clip","type":"repo","lang":"en"},{"title":"Hugging Face 博客：Vision Language Models Explained","url":"https://huggingface.co/blog/vlms","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/siglip/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/siglip/"}},{"id":"llava","name":{"zh":"视觉指令微调（LLaVA）","en":"Visual Instruction Tuning (LLaVA)"},"aka":[],"category":"multimodal","tier":"core","year":2023,"difficulty":3,"summary":{"zh":"用投影层把视觉编码器接到大模型上，再用图文指令数据微调。","en":"Connects a vision encoder to an LLM via a projector, tuned on visual instructions."},"explain":{"zh":"图像经 CLIP 编码后通过一个简单的投影层变成“视觉 token”，与文字一起输入大模型；用 GPT-4 生成的图文对话数据微调。这一“编码器+投影+LLM”范式被大量开源 VLM 采用。","en":"CLIP features pass through a simple projector into visual tokens fed with text to the LLM, tuned on GPT-4-generated visual conversations. This encoder–projector–LLM recipe is used by many open VLMs."},"paper":{"title":"Visual Instruction Tuning","authors":["Haotian Liu","Chunyuan Li","Qingyang Wu","et al."],"firstPublished":"2023-04-17","year":2023,"arxiv":"2304.08485","url":"https://arxiv.org/abs/2304.08485","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["clip","instruction-tuning"],"usedBy":[],"related":["flamingo"],"usedIn":{"zh":"LLaVA、Qwen-VL、InternVL 等开源 VLM","en":"LLaVA, Qwen-VL, InternVL and other open VLMs"},"resources":[{"title":"LLaVA 项目主页","url":"https://llava-vl.github.io/","type":"docs","lang":"en"},{"title":"Hugging Face 博客：Vision Language Models Explained","url":"https://huggingface.co/blog/vlms","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/llava/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/llava/"}},{"id":"flamingo","name":{"zh":"Flamingo 交错图文模型","en":"Flamingo"},"aka":[],"category":"multimodal","tier":"extended","year":2022,"difficulty":4,"summary":{"zh":"在冻结的大模型中插入交叉注意力层，让它读懂穿插图片的文本。","en":"Inserts cross-attention into a frozen LLM to read interleaved images and text."},"explain":{"zh":"视觉特征经 Perceiver 重采样成固定数量 token，通过新加的门控交叉注意力层注入冻结的语言模型，只需几个示例就能完成新的视觉任务。","en":"A Perceiver resampler turns vision features into a fixed set of tokens injected via new gated cross-attention layers into a frozen LM, enabling few-shot visual tasks."},"paper":{"title":"Flamingo: a Visual Language Model for Few-Shot Learning","authors":["Jean-Baptiste Alayrac","Jeff Donahue","Pauline Luc","et al."],"firstPublished":"2022-04-29","year":2022,"arxiv":"2204.14198","url":"https://arxiv.org/abs/2204.14198","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["clip","transformer"],"usedBy":[],"related":["llava"],"usedIn":{"zh":"Idefics 等开源复现","en":"Open replications such as Idefics"},"resources":[{"title":"Lilian Weng：Generalized Visual Language Models","url":"https://lilianweng.github.io/posts/2022-06-09-vlm/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/flamingo/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/flamingo/"}},{"id":"whisper","name":{"zh":"Whisper 弱监督语音识别","en":"Whisper"},"aka":[],"category":"multimodal","tier":"core","year":2022,"difficulty":2,"summary":{"zh":"用 68 万小时网络语音弱标注数据训练的通用语音识别。","en":"General speech recognition trained on 680K hours of weakly labeled web audio."},"explain":{"zh":"标准的编码器-解码器 Transformer，输入对数梅尔频谱，输出文本并用特殊 token 指定语种、任务（转写或翻译）和时间戳。靠海量多样数据获得强鲁棒性。","en":"An encoder–decoder Transformer from log-Mel spectrograms to text, with special tokens for language, task and timestamps. Massive diverse data gives robustness."},"paper":{"title":"Robust Speech Recognition via Large-Scale Weak Supervision","authors":["Alec Radford","Jong Wook Kim","Tao Xu","et al."],"firstPublished":"2022-12-06","year":2022,"arxiv":"2212.04356","url":"https://arxiv.org/abs/2212.04356","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["transformer"],"usedBy":[],"related":["vq-vae","ctc","wav2vec2"],"usedIn":{"zh":"字幕生成、会议转写、语音助手","en":"Subtitles, meeting transcription, voice assistants"},"resources":[{"title":"openai/whisper（官方实现）","url":"https://github.com/openai/whisper","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/whisper/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/whisper/"}},{"id":"sam","name":{"zh":"Segment Anything（SAM）","en":"Segment Anything (SAM)"},"aka":[],"category":"multimodal","tier":"core","year":2023,"difficulty":3,"summary":{"zh":"点一下或画个框，就能分割图中任意物体的基础模型。","en":"A promptable foundation model that segments any object from a click or box."},"explain":{"zh":"重型图像编码器只算一次，轻量的提示编码器和掩码解码器可实时响应点、框、文字等提示。配套发布了 11 亿掩码的数据集。后续 SAM 2 扩展到视频。","en":"A heavy image encoder runs once; a light prompt encoder and mask decoder respond to points, boxes and text in real time. Released with 1.1B masks; SAM 2 extends to video."},"paper":{"title":"Segment Anything","authors":["Alexander Kirillov","Eric Mintun","Nikhila Ravi","et al."],"firstPublished":"2023-04-05","year":2023,"arxiv":"2304.02643","url":"https://arxiv.org/abs/2304.02643","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["vit"],"usedBy":[],"related":["mae"],"usedIn":{"zh":"图像标注工具、医学影像、视频编辑","en":"Annotation tools, medical imaging, video editing"},"resources":[{"title":"facebookresearch/segment-anything（官方实现）","url":"https://github.com/facebookresearch/segment-anything","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/sam/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/sam/"}},{"id":"nerf-3dgs","name":{"zh":"神经辐射场与 3D 高斯泼溅","en":"NeRF & 3D Gaussian Splatting"},"aka":[],"category":"multimodal","tier":"core","year":2020,"difficulty":4,"summary":{"zh":"从多张照片重建可任意视角渲染的 3D 场景。","en":"Reconstructs 3D scenes renderable from any viewpoint from photos."},"explain":{"zh":"NeRF 用一个网络表示空间中每点的颜色和密度，通过体渲染合成新视角；3D 高斯泼溅改用数百万个可优化的 3D 高斯椭球，可实时渲染，迅速成为主流。","en":"NeRF represents color and density with a network and renders novel views volumetrically; 3D Gaussian Splatting uses millions of optimizable 3D Gaussians for real-time rendering and became mainstream."},"paper":{"title":"NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis","authors":["Ben Mildenhall","Pratul P. Srinivasan","Matthew Tancik","et al."],"firstPublished":"2020-03-19","year":2020,"arxiv":"2003.08934","url":"https://arxiv.org/abs/2003.08934","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["backprop"],"usedBy":[],"related":["world-models"],"usedIn":{"zh":"数字孪生、电商 3D、影视特效","en":"Digital twins, 3D commerce, VFX"},"resources":[{"title":"Nerfstudio 文档","url":"https://docs.nerf.studio/","type":"docs","lang":"en"},{"title":"graphdeco-inria/gaussian-splatting（官方实现）","url":"https://github.com/graphdeco-inria/gaussian-splatting","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/nerf-3dgs/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/nerf-3dgs/"}},{"id":"contrastive-learning","name":{"zh":"对比学习与 InfoNCE","en":"Contrastive Learning (InfoNCE)"},"aka":[],"category":"retrieval","tier":"core","year":2018,"difficulty":3,"summary":{"zh":"拉近正样本对、推远负样本，学到好的嵌入。","en":"Learns embeddings by pulling positives together and pushing negatives apart."},"explain":{"zh":"InfoNCE 损失把正确配对放在一堆负样本中做分类。它是 CLIP、SimCLR、句向量和检索模型的共同训练原理。","en":"InfoNCE classifies the true pair among many negatives. It is the shared objective behind CLIP, SimCLR, sentence embeddings and retrievers."},"paper":{"title":"Representation Learning with Contrastive Predictive Coding","authors":["Aaron van den Oord","Yazhe Li","Oriol Vinyals"],"firstPublished":"2018-07-10","year":2018,"arxiv":"1807.03748","url":"https://arxiv.org/abs/1807.03748","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["logistic-regression"],"usedBy":["clip","sentence-bert","dpr","simclr","wav2vec2"],"related":["clip","simclr","dpr","sentence-bert"],"usedIn":{"zh":"嵌入模型、检索模型训练","en":"Training embedding and retrieval models"},"resources":[{"title":"Lilian Weng：Contrastive Representation Learning","url":"https://lilianweng.github.io/posts/2021-05-31-contrastive/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/contrastive-learning/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/contrastive-learning/"}},{"id":"bm25","name":{"zh":"BM25 关键词检索","en":"BM25"},"aka":[],"category":"retrieval","tier":"core","year":2009,"difficulty":1,"summary":{"zh":"按词频与稀有度给文档打分的经典关键词检索。","en":"Classic keyword ranking by term frequency and rarity."},"explain":{"zh":"一个词在文档里出现越多、在全库越稀有，得分越高，并对长文档做饱和与归一化。简单却很强，在 RAG 中常与向量检索组成“混合检索”。","en":"Terms frequent in a document but rare in the corpus score high, with saturation and length normalization. Simple yet strong; paired with vectors for hybrid search in RAG."},"paper":{"title":"The Probabilistic Relevance Framework: BM25 and Beyond","authors":["Robertson et al."],"venue":"Foundations and Trends® in Information Retrieval","year":2009,"doi":"10.1561/1500000019","url":"https://doi.org/10.1561/1500000019","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":[],"usedBy":[],"related":["dpr","rag"],"usedIn":{"zh":"Elasticsearch、OpenSearch、混合检索","en":"Elasticsearch, OpenSearch, hybrid search"},"resources":[{"title":"dorianbrown/rank_bm25","url":"https://github.com/dorianbrown/rank_bm25","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/bm25/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/bm25/"}},{"id":"sentence-bert","name":{"zh":"句向量（Sentence-BERT）","en":"Sentence Embeddings (Sentence-BERT)"},"aka":[],"category":"retrieval","tier":"core","year":2019,"difficulty":2,"summary":{"zh":"把整句话编码成一个向量，用余弦相似度比较语义。","en":"Encodes sentences into vectors compared by cosine similarity."},"explain":{"zh":"用孪生 BERT 网络在句对数据上训练，让语义相近的句子向量靠近。检索时只需预先算好文档向量，查询毫秒级完成。今天的嵌入模型大多沿用这一思路。","en":"Siamese BERT trained on sentence pairs makes similar sentences close. Precompute document vectors and queries take milliseconds. Most embedding models follow this idea."},"paper":{"title":"Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks","authors":["Nils Reimers","Iryna Gurevych"],"firstPublished":"2019-08-27","year":2019,"arxiv":"1908.10084","url":"https://arxiv.org/abs/1908.10084","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["bert","contrastive-learning"],"usedBy":["dpr","matryoshka","hyde"],"related":["dpr","hnsw","matryoshka"],"usedIn":{"zh":"语义搜索、RAG 嵌入、聚类","en":"Semantic search, RAG embeddings, clustering"},"resources":[{"title":"Sentence Transformers 文档","url":"https://sbert.net/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/sentence-bert/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/sentence-bert/"}},{"id":"dpr","name":{"zh":"稠密段落检索 DPR","en":"Dense Passage Retrieval (DPR)"},"aka":[],"category":"retrieval","tier":"core","year":2020,"difficulty":3,"summary":{"zh":"用双塔编码器把问题和段落映射到同一空间进行检索。","en":"Dual encoders map questions and passages into one space for retrieval."},"explain":{"zh":"问题编码器和段落编码器分别训练，用批内负样本的对比损失；在开放域问答上大幅超过 BM25。它是 RAG 检索器的原型。","en":"Separate question and passage encoders trained with in-batch negatives beat BM25 on open-domain QA. The prototype RAG retriever."},"paper":{"title":"Dense Passage Retrieval for Open-Domain Question Answering","authors":["Vladimir Karpukhin","Barlas Oğuz","Sewon Min","et al."],"firstPublished":"2020-04-10","year":2020,"arxiv":"2004.04906","url":"https://arxiv.org/abs/2004.04906","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["sentence-bert","contrastive-learning"],"usedBy":["colbert","rag"],"related":["rag","colbert"],"usedIn":{"zh":"RAG 检索器、问答系统","en":"RAG retrievers, QA systems"},"resources":[{"title":"Sentence Transformers 文档","url":"https://sbert.net/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/dpr/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/dpr/"}},{"id":"colbert","name":{"zh":"ColBERT 后期交互","en":"ColBERT (Late Interaction)"},"aka":[],"category":"retrieval","tier":"core","year":2020,"difficulty":4,"summary":{"zh":"保留每个词的向量，在检索时做细粒度匹配，兼顾精度与速度。","en":"Keeps per-token vectors and matches them late for precise yet fast retrieval."},"explain":{"zh":"文档每个 token 都存一个向量；查询时每个查询词找最相似的文档词再求和（MaxSim）。比单向量检索更准，又比交叉编码器重排快得多，常用于重排或高精度检索。","en":"Store a vector per document token; at query time sum each query token's max similarity (MaxSim). More accurate than single vectors, much faster than cross-encoders."},"paper":{"title":"ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT","authors":["Omar Khattab","Matei Zaharia"],"firstPublished":"2020-04-27","year":2020,"arxiv":"2004.12832","url":"https://arxiv.org/abs/2004.12832","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["bert","dpr"],"usedBy":[],"related":["rag","cross-encoder-rerank"],"usedIn":{"zh":"高精度检索、重排","en":"High-precision retrieval and reranking"},"resources":[{"title":"stanford-futuredata/ColBERT（官方实现）","url":"https://github.com/stanford-futuredata/ColBERT","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/colbert/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/colbert/"}},{"id":"hnsw","name":{"zh":"HNSW 图索引","en":"HNSW (Hierarchical Navigable Small World)"},"aka":[],"category":"retrieval","tier":"core","year":2016,"difficulty":3,"summary":{"zh":"多层“小世界”图结构，实现亿级向量的毫秒级近似最近邻搜索。","en":"A layered small-world graph for millisecond approximate nearest-neighbor search."},"explain":{"zh":"上层稀疏、下层稠密，搜索时从顶层远跳快速接近目标，再逐层细化。召回高、速度快，是大多数向量数据库的默认索引。","en":"Sparse upper layers allow long jumps, denser lower layers refine. High recall and speed; the default index in most vector databases."},"paper":{"title":"Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs","authors":["Yu. A. Malkov","D. A. Yashunin"],"firstPublished":"2016-03-30","year":2016,"arxiv":"1603.09320","url":"https://arxiv.org/abs/1603.09320","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["knn"],"usedBy":["faiss"],"related":["product-quantization","faiss"],"usedIn":{"zh":"Milvus、Qdrant、pgvector、Elasticsearch 等","en":"Milvus, Qdrant, pgvector, Elasticsearch"},"resources":[{"title":"nmslib/hnswlib","url":"https://github.com/nmslib/hnswlib","type":"repo","lang":"en"},{"title":"Pinecone：HNSW 讲解","url":"https://www.pinecone.io/learn/series/faiss/hnsw/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/hnsw/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/hnsw/"}},{"id":"product-quantization","name":{"zh":"乘积量化 PQ","en":"Product Quantization (PQ)"},"aka":[],"category":"retrieval","tier":"core","year":2011,"difficulty":3,"summary":{"zh":"把向量切段、每段用小码本编码，压缩几十倍还能快速算距离。","en":"Splits vectors into chunks, each coded by a small codebook, for compressed fast search."},"explain":{"zh":"每个子向量用 k-means 码本的编号代替，一个 768 维浮点向量可压到几十字节，距离可用查表近似。常与 IVF 倒排结合处理十亿级数据。","en":"Each sub-vector is replaced by its k-means codebook index, compressing vectors to tens of bytes with table-lookup distances; combined with IVF for billion-scale search."},"paper":{"title":"Product Quantization for Nearest Neighbor Search","authors":["Jégou et al."],"venue":"IEEE Transactions on Pattern Analysis and Machine Intelligence","year":2011,"doi":"10.1109/TPAMI.2010.57","url":"https://doi.org/10.1109/TPAMI.2010.57","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["kmeans","knn"],"usedBy":["faiss"],"related":["hnsw","faiss"],"usedIn":{"zh":"FAISS IVF-PQ、大规模向量库","en":"FAISS IVF-PQ, large vector stores"},"resources":[{"title":"Pinecone：Product Quantization 讲解","url":"https://www.pinecone.io/learn/series/faiss/product-quantization/","type":"blog","lang":"en"},{"title":"FAISS Wiki","url":"https://github.com/facebookresearch/faiss/wiki","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/product-quantization/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/product-quantization/"}},{"id":"faiss","name":{"zh":"FAISS GPU 相似度搜索","en":"FAISS (Billion-scale GPU Similarity Search)"},"aka":[],"category":"retrieval","tier":"extended","year":2017,"difficulty":3,"summary":{"zh":"Meta 开源的十亿级向量检索库及其 GPU 算法。","en":"Meta's library and GPU algorithms for billion-scale vector search."},"explain":{"zh":"论文给出在 GPU 上高效实现 k 选择、IVF 与乘积量化的方法，并开源为 FAISS 库，成为向量检索事实标准工具。","en":"Describes efficient GPU k-selection, IVF and PQ, released as FAISS — the de facto vector search toolkit."},"paper":{"title":"Billion-scale similarity search with GPUs","authors":["Jeff Johnson","Matthijs Douze","Hervé Jégou"],"firstPublished":"2017-02-28","year":2017,"arxiv":"1702.08734","url":"https://arxiv.org/abs/1702.08734","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["product-quantization","hnsw"],"usedBy":[],"related":["rag"],"usedIn":{"zh":"众多向量数据库与 RAG 系统的底层","en":"Under many vector DBs and RAG systems"},"resources":[{"title":"facebookresearch/faiss","url":"https://github.com/facebookresearch/faiss","type":"repo","lang":"en"},{"title":"FAISS Wiki","url":"https://github.com/facebookresearch/faiss/wiki","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/faiss/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/faiss/"}},{"id":"matryoshka","name":{"zh":"套娃表示学习","en":"Matryoshka Representation Learning"},"aka":[],"category":"retrieval","tier":"extended","year":2022,"difficulty":3,"summary":{"zh":"一个向量的前几维本身就是可用的小向量，可按需截断。","en":"Embeddings whose prefixes are themselves usable smaller embeddings."},"explain":{"zh":"直觉：把最重要的信息塞进向量最前面的几维。训练时同时对前 8、16、32…维分别算损失，于是截断后的短向量也能用，只损失少量精度；检索时可以先用短向量粗筛、再用完整向量精排，在存储成本和精度之间灵活取舍。","en":"Intuition: pack the most important information into the first few dimensions. Training applies the loss to nested prefixes (8, 16, 32… dims), so truncated vectors still work with little loss; retrieval can shortlist with short vectors and re-rank with full ones."},"paper":{"title":"Matryoshka Representation Learning","authors":["Aditya Kusupati","Gantavya Bhatt","Aniket Rege","et al."],"firstPublished":"2022-05-26","year":2022,"arxiv":"2205.13147","url":"https://arxiv.org/abs/2205.13147","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["sentence-bert"],"usedBy":[],"related":["product-quantization"],"usedIn":{"zh":"OpenAI text-embedding-3、Sentence Transformers 的 MatryoshkaLoss 等可变维度嵌入；2026-10 的 SpecMRL（arXiv 2610.26824）用谱基重排前缀","en":"Variable-dimension embeddings such as OpenAI text-embedding-3 and Sentence Transformers MatryoshkaLoss; SpecMRL (arXiv 2610.26824, Oct 2026) reorders prefixes on a spectral basis"},"resources":[{"title":"Sentence Transformers 文档","url":"https://sbert.net/","type":"docs","lang":"en"},{"title":"Hugging Face：Introduction to Matryoshka Embedding Models","url":"https://huggingface.co/blog/matryoshka","type":"blog","lang":"en"},{"title":"RAIVNLab/MRL（官方实现）","url":"https://github.com/RAIVNLab/MRL","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/matryoshka/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/matryoshka/"}},{"id":"rag","name":{"zh":"检索增强生成 RAG","en":"Retrieval-Augmented Generation (RAG)"},"aka":[],"category":"retrieval","tier":"core","year":2020,"difficulty":2,"summary":{"zh":"先检索相关文档，再让模型基于这些资料作答。","en":"Retrieve relevant documents, then generate grounded answers."},"explain":{"zh":"把参数里的“记忆”和外部可更新的知识库结合：检索器找出相关段落，生成器把它们当上下文作答，减少幻觉、便于引用和更新知识。现在泛指一切“检索+大模型”的系统。","en":"Combine parametric memory with an updatable index: a retriever fetches passages and the generator conditions on them, reducing hallucination and enabling citations. Now refers to any retrieve-then-generate system."},"paper":{"title":"Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks","authors":["Patrick Lewis","Ethan Perez","Aleksandra Piktus","et al."],"firstPublished":"2020-05-22","year":2020,"arxiv":"2005.11401","url":"https://arxiv.org/abs/2005.11401","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["dpr","gpt3-icl"],"usedBy":["hyde","graphrag","generative-agents","prompt-injection","memgpt"],"related":["graphrag","hyde","colbert","bm25"],"usedIn":{"zh":"企业知识库问答、Perplexity 类搜索、客服","en":"Enterprise QA, Perplexity-style search, support bots"},"resources":[{"title":"Jay Alammar：The Illustrated Retrieval Transformer","url":"https://jalammar.github.io/illustrated-retrieval-transformer/","type":"blog","lang":"en"},{"title":"Hugging Face LLM 课程","url":"https://huggingface.co/learn/llm-course/chapter1/1","type":"course","lang":"multi"},{"title":"Lilian Weng：Extrinsic Hallucinations in LLMs","url":"https://lilianweng.github.io/posts/2024-07-07-hallucination/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/rag/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/rag/"}},{"id":"hyde","name":{"zh":"HyDE 假设文档检索","en":"HyDE (Hypothetical Document Embeddings)"},"aka":[],"category":"retrieval","tier":"extended","year":2022,"difficulty":2,"summary":{"zh":"先让模型写一个“假想答案”，再用它去检索真实文档。","en":"Generate a hypothetical answer, then retrieve real documents with it."},"explain":{"zh":"问题和答案的措辞常常不像。先让 LLM 生成一个可能有错的假想文档，把它编码后检索，能找到更相似的真实文档，无需标注数据。","en":"Questions and answers are phrased differently. An LLM writes a (possibly wrong) hypothetical document whose embedding retrieves real ones better, with no labeled data."},"paper":{"title":"Precise Zero-Shot Dense Retrieval without Relevance Labels","authors":["Luyu Gao","Xueguang Ma","Jimmy Lin","et al."],"firstPublished":"2022-12-20","year":2022,"arxiv":"2212.10496","url":"https://arxiv.org/abs/2212.10496","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["rag","sentence-bert"],"usedBy":[],"related":["graphrag"],"usedIn":{"zh":"RAG 查询改写","en":"Query rewriting in RAG"},"resources":[{"title":"Sentence Transformers 文档","url":"https://sbert.net/","type":"docs","lang":"en"},{"title":"texttron/hyde（HyDE 官方代码）","url":"https://github.com/texttron/hyde","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/hyde/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/hyde/"}},{"id":"graphrag","name":{"zh":"GraphRAG","en":"GraphRAG"},"aka":[],"category":"retrieval","tier":"core","year":2024,"difficulty":3,"summary":{"zh":"先从文档构建知识图谱和社区摘要，再回答全局性问题。","en":"Builds a knowledge graph and community summaries to answer global questions."},"explain":{"zh":"普通 RAG 擅长找具体片段，却答不好“整个语料的主题是什么”。GraphRAG 用 LLM 抽取实体关系、做社区聚类并逐层摘要，回答时汇总各社区摘要。","en":"Plain RAG finds snippets but struggles with corpus-wide questions. GraphRAG extracts entities and relations, clusters communities, summarizes them hierarchically, and aggregates summaries to answer."},"paper":{"title":"From Local to Global: A Graph RAG Approach to Query-Focused Summarization","authors":["Darren Edge","Ha Trinh","Newman Cheng","et al."],"firstPublished":"2024-04-24","year":2024,"arxiv":"2404.16130","url":"https://arxiv.org/abs/2404.16130","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["rag"],"usedBy":[],"related":["hyde"],"usedIn":{"zh":"Microsoft GraphRAG、企业知识分析","en":"Microsoft GraphRAG, enterprise knowledge analysis"},"resources":[{"title":"Microsoft GraphRAG 文档","url":"https://microsoft.github.io/graphrag/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/graphrag/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/graphrag/"}},{"id":"react","name":{"zh":"ReAct 推理-行动循环","en":"ReAct (Reason + Act)"},"aka":[],"category":"agents","tier":"core","year":2022,"difficulty":2,"summary":{"zh":"让模型交替“想一步、做一步、看结果”。","en":"Interleaves reasoning, actions and observations."},"explain":{"zh":"模型输出“思考”决定下一步，再输出“行动”调用搜索等工具，读取“观察”结果后继续思考，直到完成任务。这一循环是今天几乎所有 Agent 框架的骨架。","en":"The model writes a thought, then an action calling a tool, reads the observation and continues until done. This loop underlies nearly every agent framework."},"paper":{"title":"ReAct: Synergizing Reasoning and Acting in Language Models","authors":["Shunyu Yao","Jeffrey Zhao","Dian Yu","et al."],"firstPublished":"2022-10-06","year":2022,"arxiv":"2210.03629","url":"https://arxiv.org/abs/2210.03629","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["cot"],"usedBy":["function-calling","reflexion","generative-agents","swe-agent","autogen"],"related":["function-calling","reflexion","mcp"],"usedIn":{"zh":"LangChain/LangGraph、各类编码与浏览器 Agent","en":"LangChain/LangGraph, coding and browser agents"},"resources":[{"title":"ReAct 项目主页","url":"https://react-lm.github.io/","type":"docs","lang":"en"},{"title":"Lilian Weng：LLM Powered Autonomous Agents","url":"https://lilianweng.github.io/posts/2023-06-23-agent/","type":"blog","lang":"en"},{"title":"Anthropic：Building effective agents","url":"https://www.anthropic.com/engineering/building-effective-agents","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/react/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/react/"}},{"id":"toolformer","name":{"zh":"Toolformer 自学用工具","en":"Toolformer"},"aka":[],"category":"agents","tier":"extended","year":2023,"difficulty":3,"summary":{"zh":"模型自己学会何时、如何调用 API。","en":"A model teaches itself when and how to call APIs."},"explain":{"zh":"让模型在文本中插入候选 API 调用，只保留那些能让后续文本更好预测的调用作为训练数据，于是模型自监督地学会使用计算器、搜索、翻译等工具。","en":"The model inserts candidate API calls and keeps only those that improve prediction of following text, learning to use calculators, search and translation self-supervised."},"paper":{"title":"Toolformer: Language Models Can Teach Themselves to Use Tools","authors":["Timo Schick","Jane Dwivedi-Yu","Roberto Dessì","et al."],"firstPublished":"2023-02-09","year":2023,"arxiv":"2302.04761","url":"https://arxiv.org/abs/2302.04761","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["gpt3-icl"],"usedBy":[],"related":["function-calling","react"],"usedIn":{"zh":"工具学习研究","en":"Tool-learning research"},"resources":[{"title":"Lilian Weng：LLM Powered Autonomous Agents","url":"https://lilianweng.github.io/posts/2023-06-23-agent/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/toolformer/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/toolformer/"}},{"id":"function-calling","name":{"zh":"函数调用 / 工具调用","en":"Function Calling / Tool Use"},"aka":[],"category":"agents","tier":"core","year":2023,"difficulty":2,"summary":{"zh":"模型按 JSON Schema 输出结构化的函数调用参数。","en":"Models emit structured function calls that follow a JSON Schema."},"explain":{"zh":"开发者用 JSON Schema 描述可用函数，模型经专门微调后决定是否调用、输出符合签名的参数，由程序执行后把结果交回模型。OpenAI 于 2023 年 6 月在 API 中推出，现已是各家标准能力。","en":"Developers describe functions with JSON Schema; fine-tuned models decide when to call and emit matching arguments, which the app executes and returns. Launched in OpenAI's API in June 2023, now standard everywhere."},"paper":{"title":"Function calling and other API updates (OpenAI blog)","authors":["OpenAI"],"venue":"官方博客","year":2023,"url":"https://openai.com/index/function-calling-and-other-api-updates/","source":"官方页面 HTTP/WebFetch 核实","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["react"],"usedBy":["mcp","swe-agent","prompt-injection","autogen","memgpt"],"related":["mcp","toolformer"],"usedIn":{"zh":"OpenAI、Anthropic、Gemini、Qwen 等所有主流 API","en":"All major APIs: OpenAI, Anthropic, Gemini, Qwen"},"resources":[{"title":"OpenAI 文档：Function calling","url":"https://developers.openai.com/api/docs/guides/function-calling","type":"docs","lang":"en"},{"title":"Anthropic 文档：Tool use","url":"https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/function-calling/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/function-calling/"}},{"id":"mcp","name":{"zh":"模型上下文协议 MCP","en":"Model Context Protocol (MCP)"},"aka":[],"category":"agents","tier":"core","year":2024,"difficulty":2,"summary":{"zh":"连接 AI 应用与外部工具/数据源的开放标准协议。","en":"An open protocol connecting AI apps to external tools and data."},"explain":{"zh":"像 AI 的“USB-C 接口”：工具和数据源实现一次 MCP 服务器，就能被任何支持 MCP 的客户端使用。基于 JSON-RPC，定义了工具、资源和提示等能力。Anthropic 于 2024 年 11 月发布。","en":"A USB-C port for AI: implement a tool once as an MCP server and any MCP client can use it. Built on JSON-RPC with tools, resources and prompts. Released by Anthropic in Nov 2024."},"paper":{"title":"Introducing the Model Context Protocol (Anthropic)","authors":["Anthropic"],"venue":"官方公告","year":2024,"url":"https://www.anthropic.com/news/model-context-protocol","source":"官方页面 HTTP/WebFetch 核实","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["function-calling"],"usedBy":[],"related":["react"],"usedIn":{"zh":"Claude、ChatGPT、Cursor、VS Code 等客户端","en":"Clients such as Claude, ChatGPT, Cursor, VS Code"},"resources":[{"title":"Model Context Protocol 官网","url":"https://modelcontextprotocol.io/","type":"docs","lang":"en"},{"title":"MCP 规范（2025-06-18 版）","url":"https://modelcontextprotocol.io/specification/2025-06-18","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/mcp/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/mcp/"}},{"id":"reflexion","name":{"zh":"Reflexion 语言反思","en":"Reflexion"},"aka":[],"category":"agents","tier":"core","year":2023,"difficulty":3,"summary":{"zh":"失败后用文字写下反思，存进记忆，下次尝试时参考。","en":"Agents write verbal self-reflections after failures and reuse them."},"explain":{"zh":"不更新权重，而是把“这次哪里错了、下次该怎么做”的文字总结放进记忆，作为下一轮的上下文，在编程和决策任务上明显提升成功率。","en":"Instead of weight updates, store verbal lessons from failures in memory as context for the next attempt, improving coding and decision tasks."},"paper":{"title":"Reflexion: Language Agents with Verbal Reinforcement Learning","authors":["Noah Shinn","Federico Cassano","Edward Berman","et al."],"firstPublished":"2023-03-20","year":2023,"arxiv":"2303.11366","url":"https://arxiv.org/abs/2303.11366","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["react"],"usedBy":[],"related":["generative-agents","self-refine"],"usedIn":{"zh":"编码 Agent、自我修正","en":"Coding agents, self-correction"},"resources":[{"title":"noahshinn/reflexion（官方实现）","url":"https://github.com/noahshinn/reflexion","type":"repo","lang":"en"},{"title":"Lilian Weng：LLM Powered Autonomous Agents","url":"https://lilianweng.github.io/posts/2023-06-23-agent/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/reflexion/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/reflexion/"}},{"id":"generative-agents","name":{"zh":"生成式智能体与记忆","en":"Generative Agents (Memory & Reflection)"},"aka":[],"category":"agents","tier":"core","year":2023,"difficulty":3,"summary":{"zh":"带记忆流、反思和计划的 AI 角色，在小镇里模拟人类生活。","en":"Agents with memory streams, reflection and planning simulate human life."},"explain":{"zh":"每个角色把经历存成记忆流，按近因、重要性、相关性检索，并定期归纳成更高层的反思，再据此做计划。它奠定了 Agent 长期记忆的基本设计。","en":"Each agent logs experiences, retrieves by recency, importance and relevance, periodically reflects and plans — a template for long-term agent memory."},"paper":{"title":"Generative Agents: Interactive Simulacra of Human Behavior","authors":["Joon Sung Park","Joseph C. O'Brien","Carrie J. Cai","et al."],"firstPublished":"2023-04-07","year":2023,"arxiv":"2304.03442","url":"https://arxiv.org/abs/2304.03442","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["react","rag"],"usedBy":[],"related":["reflexion","memgpt","autogen"],"usedIn":{"zh":"Agent 记忆系统、社会模拟","en":"Agent memory systems, social simulation"},"resources":[{"title":"Lilian Weng：LLM Powered Autonomous Agents","url":"https://lilianweng.github.io/posts/2023-06-23-agent/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/generative-agents/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/generative-agents/"}},{"id":"swe-agent","name":{"zh":"Agent-计算机接口（SWE-agent）","en":"Agent–Computer Interface (SWE-agent)"},"aka":[],"category":"agents","tier":"core","year":2024,"difficulty":3,"summary":{"zh":"为大模型专门设计的命令行工具界面，显著提升自动修 bug 能力。","en":"LM-tailored command interfaces that boost automated bug fixing."},"explain":{"zh":"作者发现给 Agent 设计“好用的工具”（带行号的文件查看、带语法检查的编辑、精简的搜索结果）与模型能力同样重要，在 SWE-bench 上大幅提升。","en":"Designing agent-friendly tools — windowed file viewing, linted edits, concise search — matters as much as the model, greatly improving SWE-bench."},"paper":{"title":"SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering","authors":["John Yang","Carlos E. Jimenez","Alexander Wettig","et al."],"firstPublished":"2024-05-06","year":2024,"arxiv":"2405.15793","url":"https://arxiv.org/abs/2405.15793","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["react","function-calling"],"usedBy":[],"related":["mcp"],"usedIn":{"zh":"编码 Agent（Claude Code、Codex、Cursor 等同类设计）","en":"Coding agents"},"resources":[{"title":"SWE-agent/SWE-agent","url":"https://github.com/SWE-agent/SWE-agent","type":"repo","lang":"en"},{"title":"Anthropic：Building effective agents","url":"https://www.anthropic.com/engineering/building-effective-agents","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/swe-agent/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/swe-agent/"}},{"id":"q-learning","name":{"zh":"Q 学习","en":"Q-Learning"},"aka":[],"category":"rl","tier":"core","year":1992,"difficulty":2,"summary":{"zh":"学习“在某状态做某动作能拿多少长期回报”的表格或函数。","en":"Learns the long-term value of each action in each state."},"explain":{"zh":"用贝尔曼方程反复更新 Q 值：当前回报 + 折扣后的下一状态最好 Q 值。不需要环境模型，可离策略学习，是价值型强化学习的基础。","en":"Repeatedly update Q toward reward plus discounted best next-state Q (Bellman). Model-free and off-policy, it founds value-based RL."},"paper":{"title":"Q-learning","authors":["Watkins et al."],"venue":"Machine Learning","year":1992,"doi":"10.1007/BF00992698","url":"https://doi.org/10.1007/BF00992698","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["sgd"],"usedBy":["dqn","cql"],"related":["dqn"],"usedIn":{"zh":"游戏 AI、推荐、控制","en":"Game AI, recommendation, control"},"resources":[{"title":"Hugging Face 深度强化学习课程","url":"https://huggingface.co/learn/deep-rl-course/unit0/introduction","type":"course","lang":"en"},{"title":"Lilian Weng：A (Long) Peek into Reinforcement Learning","url":"https://lilianweng.github.io/posts/2018-02-19-rl-overview/","type":"blog","lang":"en"},{"title":"OpenAI Spinning Up in Deep RL","url":"https://spinningup.openai.com/en/latest/","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/q-learning/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/q-learning/"}},{"id":"dqn","name":{"zh":"深度 Q 网络 DQN","en":"Deep Q-Network (DQN)"},"aka":[],"category":"rl","tier":"core","year":2013,"difficulty":3,"summary":{"zh":"用卷积网络近似 Q 函数，直接从像素学会玩 Atari 游戏。","en":"Approximates Q with a CNN to play Atari from pixels."},"explain":{"zh":"加入经验回放（打乱相关样本）和目标网络（稳定训练目标），首次让深度网络稳定地做强化学习，在多款 Atari 游戏上达到人类水平。","en":"Experience replay and a target network stabilize training, letting deep nets do RL and reach human level on many Atari games."},"paper":{"title":"Playing Atari with Deep Reinforcement Learning","authors":["Volodymyr Mnih","Koray Kavukcuoglu","David Silver","et al."],"firstPublished":"2013-12-19","year":2013,"arxiv":"1312.5602","url":"https://arxiv.org/abs/1312.5602","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["q-learning","cnn"],"usedBy":["sac","a3c","ddpg-td3","cql","decision-transformer"],"related":["ppo","muzero"],"usedIn":{"zh":"深度强化学习入门基线","en":"Baseline for deep RL"},"resources":[{"title":"Hugging Face 深度强化学习课程","url":"https://huggingface.co/learn/deep-rl-course/unit0/introduction","type":"course","lang":"en"},{"title":"Lilian Weng：A (Long) Peek into Reinforcement Learning","url":"https://lilianweng.github.io/posts/2018-02-19-rl-overview/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/dqn/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/dqn/"}},{"id":"policy-gradient","name":{"zh":"策略梯度 REINFORCE","en":"Policy Gradient (REINFORCE)"},"aka":[],"category":"rl","tier":"core","year":1992,"difficulty":3,"summary":{"zh":"直接调整策略，让带来高回报的动作概率变大。","en":"Directly increases the probability of actions that led to high return."},"explain":{"zh":"按“动作对数概率的梯度 × 回报”更新策略参数，可处理连续动作和随机策略；方差大，常减去基线（如价值估计）。它是 PPO、GRPO 的数学基础。","en":"Update by ∇log-prob of action × return; handles continuous and stochastic policies. High variance is reduced with a baseline. It underlies PPO and GRPO."},"paper":{"title":"Simple statistical gradient-following algorithms for connectionist reinforcement learning","authors":["Williams et al."],"venue":"Machine Learning","year":1992,"doi":"10.1007/BF00992696","url":"https://doi.org/10.1007/BF00992696","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["sgd"],"usedBy":["ppo","mcts-alphazero","a3c","trpo","ddpg-td3"],"related":["ppo","grpo"],"usedIn":{"zh":"PPO/GRPO 等算法的基础","en":"Foundation of PPO and GRPO"},"resources":[{"title":"Lilian Weng：Policy Gradient Algorithms","url":"https://lilianweng.github.io/posts/2018-04-08-policy-gradient/","type":"blog","lang":"en"},{"title":"OpenAI Spinning Up in Deep RL","url":"https://spinningup.openai.com/en/latest/","type":"course","lang":"en"},{"title":"Hugging Face 深度强化学习课程","url":"https://huggingface.co/learn/deep-rl-course/unit0/introduction","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/policy-gradient/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/policy-gradient/"}},{"id":"ppo","name":{"zh":"近端策略优化 PPO","en":"Proximal Policy Optimization (PPO)"},"aka":[],"category":"rl","tier":"core","year":2017,"difficulty":3,"summary":{"zh":"限制每次策略更新幅度，稳定又简单的策略梯度算法。","en":"A stable, simple policy-gradient method that clips update size."},"explain":{"zh":"用“新旧策略概率比”乘优势估计做目标，并把比值裁剪在 1±ε 内，防止一步走太远。易实现、效果稳，是机器人、游戏和 RLHF 的默认算法。","en":"Optimize ratio × advantage with the ratio clipped to 1±ε to avoid destructive steps. Easy and stable — the default for robotics, games and RLHF."},"paper":{"title":"Proximal Policy Optimization Algorithms","authors":["John Schulman","Filip Wolski","Prafulla Dhariwal","et al."],"firstPublished":"2017-07-20","year":2017,"arxiv":"1707.06347","url":"https://arxiv.org/abs/1707.06347","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["policy-gradient"],"usedBy":["rlhf","grpo","rlvr","sac"],"related":["rlhf","grpo","sac","trpo","a3c"],"usedIn":{"zh":"RLHF、OpenAI Five、机器人控制","en":"RLHF, OpenAI Five, robot control"},"resources":[{"title":"Spinning Up：PPO","url":"https://spinningup.openai.com/en/latest/algorithms/ppo.html","type":"docs","lang":"en"},{"title":"TRL：PPO Trainer","url":"https://huggingface.co/docs/trl/ppo_trainer","type":"docs","lang":"en"},{"title":"Lilian Weng：Policy Gradient Algorithms","url":"https://lilianweng.github.io/posts/2018-04-08-policy-gradient/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/ppo/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/ppo/"}},{"id":"sac","name":{"zh":"软演员-评论家 SAC","en":"Soft Actor-Critic (SAC)"},"aka":[],"category":"rl","tier":"extended","year":2018,"difficulty":4,"summary":{"zh":"在最大化回报的同时最大化策略熵，样本效率高的离策略算法。","en":"Off-policy actor-critic maximizing both return and entropy."},"explain":{"zh":"鼓励策略保持随机以充分探索，配合经验回放和双 Q 网络，在连续控制任务上又稳又省样本，是机器人学习常用算法。","en":"Encourages exploration via an entropy bonus, with replay and twin Q networks — stable and sample-efficient for continuous control."},"paper":{"title":"Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor","authors":["Tuomas Haarnoja","Aurick Zhou","Pieter Abbeel","et al."],"firstPublished":"2018-01-04","year":2018,"arxiv":"1801.01290","url":"https://arxiv.org/abs/1801.01290","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["ppo","dqn"],"usedBy":["dreamer"],"related":["dreamer"],"usedIn":{"zh":"机器人控制、连续控制","en":"Robot and continuous control"},"resources":[{"title":"Spinning Up：SAC","url":"https://spinningup.openai.com/en/latest/algorithms/sac.html","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/sac/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/sac/"}},{"id":"mcts-alphazero","name":{"zh":"AlphaGo / AlphaZero（自对弈 + 蒙特卡洛树搜索）","en":"AlphaGo / AlphaZero (Self-Play + MCTS)"},"aka":[],"category":"rl","tier":"core","year":2016,"difficulty":4,"summary":{"zh":"神经网络评估局面、树搜索向前推演，靠自我对弈从零学会下棋。","en":"Neural evaluation plus tree search, learned from scratch by self-play."},"explain":{"zh":"策略网络提出候选走法、价值网络评估局面，蒙特卡洛树搜索在两者指导下推演；自我对弈产生的数据再训练网络。AlphaZero 不用人类棋谱就在围棋、国际象棋、将棋上超越人类。","en":"A policy net proposes moves, a value net evaluates, MCTS searches with both, and self-play data retrains the nets. AlphaZero surpassed humans in Go, chess and shogi without human games."},"paper":{"title":"Mastering the game of Go with deep neural networks and tree search","authors":["Silver et al."],"venue":"Nature","year":2016,"doi":"10.1038/nature16961","url":"https://doi.org/10.1038/nature16961","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["policy-gradient","cnn"],"usedBy":["muzero"],"related":["muzero","tree-of-thoughts","test-time-scaling"],"usedIn":{"zh":"棋类 AI、Leela Chess Zero、KataGo","en":"Game AIs: Leela Chess Zero, KataGo"},"resources":[{"title":"Google DeepMind：AlphaZero","url":"https://deepmind.google/blog/alphazero-shedding-new-light-on-chess-shogi-and-go/","type":"blog","lang":"en"},{"title":"Hugging Face 深度强化学习课程","url":"https://huggingface.co/learn/deep-rl-course/unit0/introduction","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/mcts-alphazero/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/mcts-alphazero/"}},{"id":"muzero","name":{"zh":"MuZero","en":"MuZero"},"aka":[],"category":"rl","tier":"extended","year":2019,"difficulty":5,"summary":{"zh":"不知道规则也能规划：学一个只预测有用量的内部模型。","en":"Plans without knowing the rules by learning a value-focused model."},"explain":{"zh":"在 AlphaZero 基础上，学习一个隐空间动力学模型，只预测奖励、价值和策略，而不重建画面，在 Atari 和棋类上都达到顶尖水平。","en":"Extends AlphaZero with a learned latent dynamics model predicting only reward, value and policy, reaching top performance in Atari and board games."},"paper":{"title":"Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model","authors":["Julian Schrittwieser","Ioannis Antonoglou","Thomas Hubert","et al."],"firstPublished":"2019-11-19","year":2019,"arxiv":"1911.08265","url":"https://arxiv.org/abs/1911.08265","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["mcts-alphazero"],"usedBy":[],"related":["dreamer","world-models"],"usedIn":{"zh":"视频压缩码率控制等","en":"Applied to e.g. video-compression rate control"},"resources":[{"title":"Google DeepMind：MuZero","url":"https://deepmind.google/blog/muzero-mastering-go-chess-shogi-and-atari-without-rules/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/muzero/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/muzero/"}},{"id":"world-models","name":{"zh":"世界模型（Ha & Schmidhuber）","en":"World Models"},"aka":[],"category":"world-models","tier":"core","year":2018,"difficulty":3,"summary":{"zh":"先学一个能“做梦”的环境模型，再在梦里训练策略。","en":"Learn a dreamable environment model, then train the policy inside it."},"explain":{"zh":"VAE 把画面压成隐向量，RNN 预测下一刻隐状态，一个很小的控制器在这个“梦境”里学会驾驶和躲避。它重新点燃了世界模型研究。","en":"A VAE compresses frames, an RNN predicts the next latent, and a tiny controller learns inside the dream. It revived world-model research."},"paper":{"title":"World Models","authors":["David Ha","Jürgen Schmidhuber"],"firstPublished":"2018-03-27","year":2018,"arxiv":"1803.10122","url":"https://arxiv.org/abs/1803.10122","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["vae","lstm"],"usedBy":["dreamer","genie"],"related":["dreamer","genie","jepa"],"usedIn":{"zh":"世界模型研究的起点","en":"Starting point for world-model research"},"resources":[{"title":"World Models 交互论文主页","url":"https://worldmodels.github.io/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/world-models/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/world-models/"}},{"id":"dreamer","name":{"zh":"Dreamer（隐空间想象）","en":"Dreamer (Latent Imagination)"},"aka":[],"category":"world-models","tier":"core","year":2019,"difficulty":4,"summary":{"zh":"在学到的隐空间世界模型中想象未来轨迹来训练行为。","en":"Trains behavior by imagining trajectories in a learned latent world model."},"explain":{"zh":"学习循环状态空间模型预测隐状态与奖励，演员-评论家完全在想象轨迹上学习。DreamerV3 用一套固定超参解决 150 多个任务，并首次从零在 Minecraft 中挖到钻石。","en":"Learn a recurrent state-space model; train actor–critic on imagined rollouts. DreamerV3 solves 150+ tasks with fixed hyperparameters and first collected diamonds in Minecraft from scratch."},"paper":{"title":"Dream to Control: Learning Behaviors by Latent Imagination","authors":["Danijar Hafner","Timothy Lillicrap","Jimmy Ba","et al."],"firstPublished":"2019-12-03","year":2019,"arxiv":"1912.01603","url":"https://arxiv.org/abs/1912.01603","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["world-models","sac"],"usedBy":[],"related":["muzero","genie"],"usedIn":{"zh":"机器人与游戏中的模型式强化学习","en":"Model-based RL in robotics and games"},"resources":[{"title":"danijar/dreamerv3（官方实现）","url":"https://github.com/danijar/dreamerv3","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/dreamer/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/dreamer/"}},{"id":"jepa","name":{"zh":"联合嵌入预测架构 JEPA","en":"Joint-Embedding Predictive Architecture (JEPA)"},"aka":[],"category":"world-models","tier":"core","year":2023,"difficulty":4,"summary":{"zh":"在抽象表征空间里预测被遮住的部分，而不是重建像素。","en":"Predicts masked content in representation space rather than pixels."},"explain":{"zh":"LeCun 倡导的路线：用上下文块的表征去预测目标块的表征，忽略不可预测的细节，学到更语义化的特征。I-JEPA 用于图像，V-JEPA 用于视频。","en":"LeCun's approach: predict target-block representations from context-block representations, ignoring unpredictable detail. I-JEPA for images, V-JEPA for video."},"paper":{"title":"Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture","authors":["Mahmoud Assran","Quentin Duval","Ishan Misra","et al."],"firstPublished":"2023-01-19","year":2023,"arxiv":"2301.08243","url":"https://arxiv.org/abs/2301.08243","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["mae","vit"],"usedBy":[],"related":["world-models"],"usedIn":{"zh":"Meta I-JEPA / V-JEPA / V-JEPA 2（视频理解、预测与机器人规划）；2026-09 的 PDE-JEPA（arXiv 2609.34715）把它用于偏微分方程动力学","en":"Meta I-JEPA / V-JEPA / V-JEPA 2 (video understanding, prediction and robot planning); PDE-JEPA (arXiv 2609.34715, Sep 2026) applies it to PDE dynamics"},"resources":[{"title":"facebookresearch/ijepa（官方实现）","url":"https://github.com/facebookresearch/ijepa","type":"repo","lang":"en"},{"title":"facebookresearch/vjepa2（V-JEPA 2 官方实现）","url":"https://github.com/facebookresearch/vjepa2","type":"repo","lang":"en"},{"title":"Meta AI：I-JEPA 介绍博客","url":"https://ai.meta.com/blog/yann-lecun-ai-model-i-jepa/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/jepa/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/jepa/"}},{"id":"genie","name":{"zh":"Genie 可交互世界生成","en":"Genie (Generative Interactive Environments)"},"aka":[],"category":"world-models","tier":"core","year":2024,"difficulty":5,"summary":{"zh":"从无标注游戏视频中学会生成可用键盘操控的世界。","en":"Learns playable, controllable worlds from unlabeled videos."},"explain":{"zh":"由视频分词器、隐动作模型和动力学模型组成，无需动作标签就推断出“玩家做了什么”，于是一张图片就能变成可交互的 2D 世界；后续 Genie 2/3 扩展到 3D。","en":"A video tokenizer, latent action model and dynamics model infer actions without labels, turning an image into a playable 2D world; Genie 2/3 extend to 3D."},"paper":{"title":"Genie: Generative Interactive Environments","authors":["Jake Bruce","Michael Dennis","Ashley Edwards","et al."],"firstPublished":"2024-02-23","year":2024,"arxiv":"2402.15391","url":"https://arxiv.org/abs/2402.15391","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["vq-vae","transformer","world-models"],"usedBy":[],"related":["dit"],"usedIn":{"zh":"Genie 2/3、游戏与机器人训练环境","en":"Genie 2/3, game and robotics training environments"},"resources":[{"title":"Google DeepMind：Genie 2","url":"https://deepmind.google/blog/genie-2-a-large-scale-foundation-world-model/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/genie/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/genie/"}},{"id":"simclr","name":{"zh":"SimCLR 自监督对比学习","en":"SimCLR"},"aka":[],"category":"world-models","tier":"extended","year":2020,"difficulty":3,"summary":{"zh":"同一图片的两种增强视图互为正样本，无标签学习视觉表征。","en":"Learns visual features without labels by matching two augmentations."},"explain":{"zh":"对同一图像做两次随机裁剪、调色等增强，让它们的表示靠近、与其他图片远离。证明了强数据增强和大 batch 的重要性。","en":"Two random augmentations of an image are pulled together and pushed from others, showing the importance of strong augmentation and large batches."},"paper":{"title":"A Simple Framework for Contrastive Learning of Visual Representations","authors":["Ting Chen","Simon Kornblith","Mohammad Norouzi","et al."],"firstPublished":"2020-02-13","year":2020,"arxiv":"2002.05709","url":"https://arxiv.org/abs/2002.05709","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["contrastive-learning","resnet"],"usedBy":["dinov2"],"related":["mae","clip"],"usedIn":{"zh":"视觉预训练","en":"Visual pretraining"},"resources":[{"title":"google-research/simclr","url":"https://github.com/google-research/simclr","type":"repo","lang":"en"},{"title":"Lilian Weng：Contrastive Representation Learning","url":"https://lilianweng.github.io/posts/2021-05-31-contrastive/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/simclr/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/simclr/"}},{"id":"mae","name":{"zh":"掩码自编码器 MAE","en":"Masked Autoencoders (MAE)"},"aka":[],"category":"world-models","tier":"core","year":2021,"difficulty":3,"summary":{"zh":"遮住 75% 的图像块，让模型补全，从而学到视觉表征。","en":"Masks 75% of image patches and reconstructs them to learn representations."},"explain":{"zh":"像 BERT 的完形填空搬到视觉：编码器只看可见块，轻量解码器重建被遮块的像素。高遮盖率让任务足够难，训练又快。","en":"BERT-style masking for vision: the encoder sees only visible patches and a light decoder reconstructs the rest. High masking makes the task hard yet training fast."},"paper":{"title":"Masked Autoencoders Are Scalable Vision Learners","authors":["Kaiming He","Xinlei Chen","Saining Xie","et al."],"firstPublished":"2021-11-11","year":2021,"arxiv":"2111.06377","url":"https://arxiv.org/abs/2111.06377","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["vit","bert"],"usedBy":["jepa"],"related":["jepa","simclr"],"usedIn":{"zh":"视觉预训练、医学影像","en":"Visual pretraining, medical imaging"},"resources":[{"title":"facebookresearch/mae","url":"https://github.com/facebookresearch/mae","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/mae/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/mae/"}},{"id":"gnn","name":{"zh":"图神经网络（GCN / GAT）","en":"Graph Neural Networks (GCN / GAT)"},"aka":[],"category":"graph-science","tier":"core","year":2016,"difficulty":3,"summary":{"zh":"每个节点聚合邻居信息来更新自己，用于图结构数据。","en":"Nodes update by aggregating their neighbors' features."},"explain":{"zh":"GCN 用归一化的邻居平均做“消息传递”，GAT 用注意力给不同邻居不同权重。广泛用于分子、社交网络、推荐和交通预测。","en":"GCN averages normalized neighbor messages; GAT weights neighbors with attention. Used for molecules, social networks, recommendation and traffic."},"paper":{"title":"Semi-Supervised Classification with Graph Convolutional Networks","authors":["Thomas N. Kipf","Max Welling"],"firstPublished":"2016-09-09","year":2016,"arxiv":"1609.02907","url":"https://arxiv.org/abs/1609.02907","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["backprop"],"usedBy":["alphafold"],"related":["alphafold","transformer"],"usedIn":{"zh":"药物发现、推荐系统、知识图谱","en":"Drug discovery, recommendation, knowledge graphs"},"resources":[{"title":"Distill：A Gentle Introduction to Graph Neural Networks","url":"https://distill.pub/2021/gnn-intro/","type":"blog","lang":"en"},{"title":"PyTorch Geometric 文档","url":"https://pytorch-geometric.readthedocs.io/en/latest/","type":"docs","lang":"en"},{"title":"Stanford CS224N 深度学习自然语言处理","url":"https://web.stanford.edu/class/cs224n/","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/gnn/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/gnn/"}},{"id":"alphafold","name":{"zh":"AlphaFold 2 蛋白质结构预测","en":"AlphaFold 2"},"aka":[],"category":"graph-science","tier":"core","year":2021,"difficulty":4,"summary":{"zh":"从氨基酸序列预测蛋白质 3D 结构，达到实验精度。","en":"Predicts protein 3D structure from sequence at near-experimental accuracy."},"explain":{"zh":"Evoformer 模块在多序列比对和残基对表示之间反复交换信息，结构模块用不变点注意力直接输出原子坐标。它解决了 50 年的蛋白质折叠难题，获 2024 年诺贝尔化学奖。","en":"The Evoformer exchanges information between MSA and pair representations; a structure module with invariant point attention outputs atom coordinates. It cracked protein folding and shared the 2024 Chemistry Nobel."},"paper":{"title":"Highly accurate protein structure prediction with AlphaFold","authors":["Jumper et al."],"venue":"Nature","year":2021,"doi":"10.1038/s41586-021-03819-2","url":"https://doi.org/10.1038/s41586-021-03819-2","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["transformer","gnn"],"usedBy":[],"related":["alphageometry"],"usedIn":{"zh":"结构生物学、药物设计；AlphaFold 3","en":"Structural biology, drug design; AlphaFold 3"},"resources":[{"title":"google-deepmind/alphafold","url":"https://github.com/google-deepmind/alphafold","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/alphafold/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/alphafold/"}},{"id":"alphageometry","name":{"zh":"AlphaGeometry 神经-符号推理","en":"AlphaGeometry (Neuro-Symbolic Reasoning)"},"aka":[],"category":"graph-science","tier":"core","year":2024,"difficulty":4,"summary":{"zh":"语言模型提出辅助线，符号引擎严格推导，解奥数几何题。","en":"An LM proposes constructions; a symbolic engine proves olympiad geometry."},"explain":{"zh":"用 1 亿条合成定理和证明训练语言模型，在符号推理卡住时提出辅助构造，符号引擎负责可验证的推导，在 IMO 几何题上接近金牌选手水平。","en":"Trained on 100M synthetic theorems, the LM suggests auxiliary constructions when the symbolic engine stalls; the engine produces verifiable proofs, approaching IMO gold-medalist level on geometry."},"paper":{"title":"Solving olympiad geometry without human demonstrations","authors":["Trinh et al."],"venue":"Nature","year":2024,"doi":"10.1038/s41586-023-06747-5","url":"https://doi.org/10.1038/s41586-023-06747-5","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["transformer"],"usedBy":[],"related":["alphafold","rlvr"],"usedIn":{"zh":"数学推理、形式化证明","en":"Math reasoning, formal proof"},"resources":[{"title":"google-deepmind/alphageometry","url":"https://github.com/google-deepmind/alphageometry","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/alphageometry/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/alphageometry/"}},{"id":"attribution","name":{"zh":"特征归因（积分梯度 / SHAP）","en":"Feature Attribution (Integrated Gradients / SHAP)"},"aka":[],"category":"interpretability","tier":"core","year":2017,"difficulty":2,"summary":{"zh":"计算每个输入特征对预测贡献了多少。","en":"Measures how much each input feature contributed to a prediction."},"explain":{"zh":"积分梯度沿从“空白基线”到真实输入的路径累积梯度；SHAP 用博弈论的 Shapley 值公平分配贡献。适合解释表格模型和分类器为何做出某个判断。","en":"Integrated Gradients accumulates gradients along a path from a baseline to the input; SHAP fairly splits credit with Shapley values. Good for explaining tabular models and classifiers."},"paper":{"title":"Axiomatic Attribution for Deep Networks","authors":["Mukund Sundararajan","Ankur Taly","Qiqi Yan"],"firstPublished":"2017-03-04","year":2017,"arxiv":"1703.01365","url":"https://arxiv.org/abs/1703.01365","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["backprop"],"usedBy":[],"related":["grad-cam","interp-sae"],"usedIn":{"zh":"金融风控解释、医疗模型审计","en":"Credit-model explanations, medical model audits"},"resources":[{"title":"shap/shap","url":"https://github.com/shap/shap","type":"repo","lang":"en"},{"title":"Captum（PyTorch 可解释性库）","url":"https://captum.ai/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/attribution/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/attribution/"}},{"id":"grad-cam","name":{"zh":"Grad-CAM 热力图","en":"Grad-CAM"},"aka":[],"category":"interpretability","tier":"extended","year":2016,"difficulty":2,"summary":{"zh":"用梯度生成热力图，显示 CNN 看图时关注哪里。","en":"Gradient-weighted heatmaps showing where a CNN looks."},"explain":{"zh":"用目标类别对最后一个卷积层特征图的梯度给通道加权求和，得到粗粒度的定位热力图，帮助发现模型是否“看错地方”。","en":"Weight the last conv layer's channels by the class gradient to get a coarse localization heatmap, revealing whether the model looks at the right place."},"paper":{"title":"Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization","authors":["Ramprasaath R. Selvaraju","Michael Cogswell","Abhishek Das","et al."],"firstPublished":"2016-10-07","year":2016,"arxiv":"1610.02391","url":"https://arxiv.org/abs/1610.02391","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["cnn","backprop"],"usedBy":[],"related":["attribution"],"usedIn":{"zh":"医学影像与视觉模型调试","en":"Debugging medical and vision models"},"resources":[{"title":"jacobgil/pytorch-grad-cam","url":"https://github.com/jacobgil/pytorch-grad-cam","type":"repo","lang":"en"},{"title":"Captum（PyTorch 可解释性库）","url":"https://captum.ai/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/grad-cam/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/grad-cam/"}},{"id":"logit-lens","name":{"zh":"Logit Lens 与探针","en":"Logit Lens & Probing"},"aka":[],"category":"interpretability","tier":"core","year":2020,"difficulty":3,"summary":{"zh":"把中间层的表示直接解码成词，或用小分类器探测其中编码的信息。","en":"Decodes intermediate layers into tokens, or probes them with small classifiers."},"explain":{"zh":"Logit Lens 把每层的隐藏状态乘以输出矩阵，看模型在中途“想到了”哪个词；探针则训练线性分类器检查某层是否编码了语法、真假等信息。它们是理解模型内部的入门工具。","en":"The logit lens multiplies each layer's hidden state by the unembedding to see the model's interim guess; probes train linear classifiers to test what a layer encodes. Entry-level tools for model internals."},"paper":{"title":"interpreting GPT: the logit lens","authors":["nostalgebraist"],"venue":"LessWrong","year":2020,"url":"https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens","source":"官方页面 HTTP/WebFetch 核实","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["transformer"],"usedBy":["induction-heads","rome","activation-steering"],"related":["induction-heads","activation-steering"],"usedIn":{"zh":"可解释性研究入门","en":"Intro interpretability research"},"resources":[{"title":"nostalgebraist：interpreting GPT: the logit lens","url":"https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens","type":"blog","lang":"en"},{"title":"TransformerLens","url":"https://github.com/TransformerLensOrg/TransformerLens","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/logit-lens/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/logit-lens/"}},{"id":"induction-heads","name":{"zh":"电路与归纳头","en":"Circuits & Induction Heads"},"aka":[],"category":"interpretability","tier":"core","year":2022,"difficulty":4,"summary":{"zh":"找出模型中实现具体算法的“电路”，如复制前文模式的归纳头。","en":"Finds circuits implementing algorithms, like induction heads that copy patterns."},"explain":{"zh":"Anthropic 在小型 Transformer 中发现两层注意力头组成的“归纳头”：看到 A B … A 时预测 B。它们在训练中突然出现，并与上下文学习能力同时形成。","en":"Two attention heads form an induction head: after A B … A, predict B. They appear abruptly during training, coinciding with in-context learning."},"paper":{"title":"In-context Learning and Induction Heads","authors":["Catherine Olsson","Nelson Elhage","Neel Nanda","et al."],"firstPublished":"2022-09-24","year":2022,"arxiv":"2209.11895","url":"https://arxiv.org/abs/2209.11895","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["transformer","logit-lens"],"usedBy":["circuit-tracing"],"related":["superposition","circuit-tracing"],"usedIn":{"zh":"机制可解释性研究","en":"Mechanistic interpretability research"},"resources":[{"title":"Anthropic：A Mathematical Framework for Transformer Circuits","url":"https://transformer-circuits.pub/2021/framework/index.html","type":"paper","lang":"en"},{"title":"Transformer Circuits Thread","url":"https://transformer-circuits.pub/","type":"blog","lang":"en"},{"title":"TransformerLens","url":"https://github.com/TransformerLensOrg/TransformerLens","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/induction-heads/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/induction-heads/"}},{"id":"superposition","name":{"zh":"叠加假说","en":"Superposition"},"aka":[],"category":"interpretability","tier":"core","year":2022,"difficulty":4,"summary":{"zh":"模型用比神经元更多的“特征”，把它们叠加压缩在同一组维度里。","en":"Models pack more features than dimensions by superimposing them."},"explain":{"zh":"玩具模型实验表明，当特征稀疏时，网络会把多个特征编码在近似正交的方向上，导致单个神经元“一神多职”（多义）。这解释了为何逐神经元解读困难，也催生了稀疏自编码器。","en":"Toy models show that sparse features get stored in nearly orthogonal directions, making neurons polysemantic. This explains why neuron-level reading fails and motivated sparse autoencoders."},"paper":{"title":"Toy Models of Superposition","authors":["Nelson Elhage","Tristan Hume","Catherine Olsson","et al."],"firstPublished":"2022-09-21","year":2022,"arxiv":"2209.10652","url":"https://arxiv.org/abs/2209.10652","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["pca"],"usedBy":["interp-sae"],"related":["interp-sae"],"usedIn":{"zh":"可解释性理论基础","en":"Theory underpinning interpretability"},"resources":[{"title":"Transformer Circuits Thread","url":"https://transformer-circuits.pub/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/superposition/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/superposition/"}},{"id":"interp-sae","name":{"zh":"稀疏自编码器 SAE","en":"Sparse Autoencoders (SAE)"},"aka":[],"category":"interpretability","tier":"core","year":2023,"difficulty":4,"summary":{"zh":"把模型激活分解成大量可解释的稀疏“特征”。","en":"Decomposes activations into many sparse, interpretable features."},"explain":{"zh":"训练一个比激活维度宽得多、但每次只激活少数单元的自编码器重建模型内部激活，每个单元往往对应一个清晰概念（如“金门大桥”“代码漏洞”）。Anthropic、OpenAI、DeepMind 均已扩展到前沿模型。","en":"A much wider autoencoder with sparse activations reconstructs internal activations; units often map to crisp concepts (Golden Gate Bridge, code bugs). Anthropic, OpenAI and DeepMind scaled it to frontier models."},"paper":{"title":"Sparse Autoencoders Find Highly Interpretable Features in Language Models","authors":["Hoagy Cunningham","Aidan Ewart","Logan Riggs","et al."],"firstPublished":"2023-09-15","year":2023,"arxiv":"2309.08600","url":"https://arxiv.org/abs/2309.08600","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["superposition","vae"],"usedBy":["circuit-tracing"],"related":["circuit-tracing","activation-steering"],"usedIn":{"zh":"Claude 特征研究、Gemma Scope、Neuronpedia","en":"Claude features, Gemma Scope, Neuronpedia"},"resources":[{"title":"Anthropic：Towards Monosemanticity","url":"https://transformer-circuits.pub/2023/monosemantic-features","type":"paper","lang":"en"},{"title":"Anthropic：Scaling Monosemanticity","url":"https://transformer-circuits.pub/2024/scaling-monosemanticity/","type":"paper","lang":"en"},{"title":"SAELens","url":"https://github.com/decoderesearch/SAELens","type":"repo","lang":"en"},{"title":"Neuronpedia","url":"https://www.neuronpedia.org/","type":"tool","lang":"en"},{"title":"Google DeepMind：Gemma Scope","url":"https://deepmind.google/blog/gemma-scope-helping-the-safety-community-shed-light-on-the-inner-workings-of-language-models/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/interp-sae/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/interp-sae/"}},{"id":"circuit-tracing","name":{"zh":"电路追踪与归因图","en":"Circuit Tracing & Attribution Graphs"},"aka":[],"category":"interpretability","tier":"core","year":2025,"difficulty":5,"summary":{"zh":"用替代模型画出某次回答中特征之间的因果路径图。","en":"Maps causal feature pathways behind a specific response."},"explain":{"zh":"用跨层转码器替换模型 MLP，得到可解释的特征，再为一次具体的输出构建归因图，看到模型如何多步推理、提前规划押韵等。Anthropic 2025 年发布并开源工具。","en":"Replace MLPs with cross-layer transcoders, then build attribution graphs for a specific output, revealing multi-step reasoning and planning ahead. Published and open-sourced by Anthropic in 2025."},"paper":{"title":"Circuit Tracing: Revealing Computational Graphs in Language Models","authors":["Ameisen et al. (Anthropic)"],"venue":"Transformer Circuits Thread","year":2025,"url":"https://transformer-circuits.pub/2025/attribution-graphs/methods.html","source":"官方页面 HTTP/WebFetch 核实","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["interp-sae","induction-heads"],"usedBy":[],"related":["activation-steering"],"usedIn":{"zh":"Claude 3.5 Haiku 机制研究","en":"Mechanism studies of Claude 3.5 Haiku"},"resources":[{"title":"Anthropic：Circuit Tracing（方法论文）","url":"https://transformer-circuits.pub/2025/attribution-graphs/methods.html","type":"paper","lang":"en"},{"title":"Neuronpedia","url":"https://www.neuronpedia.org/","type":"tool","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/circuit-tracing/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/circuit-tracing/"}},{"id":"rome","name":{"zh":"知识定位与编辑 ROME","en":"Locating & Editing Facts (ROME)"},"aka":[],"category":"interpretability","tier":"extended","year":2022,"difficulty":4,"summary":{"zh":"找到事实存储在哪一层，并直接修改那里的权重。","en":"Locates where facts are stored and edits those weights directly."},"explain":{"zh":"用因果追踪发现事实知识主要由中层 MLP 在主语最后一个 token 处调取，然后对该 MLP 做一次秩一更新，即可改写某条事实（如“埃菲尔铁塔在罗马”）。","en":"Causal tracing shows facts are recalled by mid-layer MLPs at the subject's last token; a rank-one update to that MLP rewrites a fact."},"paper":{"title":"Locating and Editing Factual Associations in GPT","authors":["Kevin Meng","David Bau","Alex Andonian","et al."],"firstPublished":"2022-02-10","year":2022,"arxiv":"2202.05262","url":"https://arxiv.org/abs/2202.05262","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["transformer","logit-lens"],"usedBy":[],"related":["activation-steering"],"usedIn":{"zh":"知识编辑研究","en":"Knowledge editing research"},"resources":[{"title":"ROME 项目主页","url":"https://rome.baulab.info/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/rome/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/rome/"}},{"id":"activation-steering","name":{"zh":"激活引导","en":"Activation Steering"},"aka":["Abliteration","拒绝方向消融","Directional Ablation","Steering Vector"],"category":"interpretability","tier":"core","year":2023,"difficulty":3,"summary":{"zh":"在推理时给中间激活加一个“方向向量”来改变模型行为。","en":"Adds a direction vector to activations at inference to change behavior."},"explain":{"zh":"直觉：模型内部的某些“方向”对应某种行为。用两组对比提示（如“爱”与“恨”）的激活差得到引导向量，推理时加到某层，就能让模型更积极、更诚实或谈论特定话题，无需微调。反过来，把“拒绝”方向从权重里投影掉（俗称 abliteration），模型就不再拒答——这也说明仅靠微调式安全对齐很容易被移除。","en":"Intuition: certain directions inside a model map to behaviours. The activation difference between contrasting prompts gives a steering vector that, added at a layer, makes outputs more positive, honest or on-topic without fine-tuning. Projecting the “refusal” direction out of the weights (“abliteration”) removes refusals, showing how easily fine-tuned safety can be stripped."},"paper":{"title":"Steering Language Models With Activation Engineering","authors":["Alexander Matt Turner","Lisa Thiergart","Gavin Leech","et al."],"firstPublished":"2023-08-20","year":2023,"arxiv":"2308.10248","url":"https://arxiv.org/abs/2308.10248","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["logit-lens"],"usedBy":[],"related":["interp-sae","rome"],"usedIn":{"zh":"行为控制与安全研究；社区据“拒绝方向”制作去审查（abliterated）开源模型，2026-10 德国 t3n 报道了其安全风险","en":"Behaviour control and safety research; the community uses the refusal direction to make “abliterated” uncensored open models, whose risks t3n reported on in Oct 2026"},"resources":[{"title":"TransformerLens","url":"https://github.com/TransformerLensOrg/TransformerLens","type":"repo","lang":"en"},{"title":"andyrdt/refusal_direction（Arditi 等“拒绝方向”论文官方代码）","url":"https://github.com/andyrdt/refusal_direction","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/activation-steering/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/activation-steering/"}},{"id":"adversarial-examples","name":{"zh":"对抗样本（FGSM）","en":"Adversarial Examples (FGSM)"},"aka":[],"category":"safety","tier":"core","year":2014,"difficulty":2,"summary":{"zh":"在输入上加人眼几乎看不出的扰动，就能骗过模型。","en":"Imperceptible input perturbations that fool models."},"explain":{"zh":"沿损失梯度的符号方向给图片加极小扰动，模型就会高置信度地误判。作者认为根源是模型的高维线性行为，并提出对抗训练作为防御。","en":"A tiny step along the sign of the loss gradient makes models misclassify confidently. The authors blame high-dimensional linearity and propose adversarial training."},"paper":{"title":"Explaining and Harnessing Adversarial Examples","authors":["Ian J. Goodfellow","Jonathon Shlens","Christian Szegedy"],"firstPublished":"2014-12-20","year":2014,"arxiv":"1412.6572","url":"https://arxiv.org/abs/1412.6572","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["backprop","cnn"],"usedBy":["gcg-jailbreak","sleeper-agents"],"related":["gcg-jailbreak"],"usedIn":{"zh":"自动驾驶、人脸识别的鲁棒性测试","en":"Robustness testing for autonomous driving and face recognition"},"resources":[{"title":"Stanford CS231n 视觉识别中的卷积网络","url":"https://cs231n.github.io/","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/adversarial-examples/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/adversarial-examples/"}},{"id":"gcg-jailbreak","name":{"zh":"GCG 通用越狱攻击","en":"GCG Universal Jailbreak Attack"},"aka":[],"category":"safety","tier":"core","year":2023,"difficulty":3,"summary":{"zh":"自动搜索一段“乱码后缀”，让对齐过的模型输出有害内容。","en":"Automatically finds gibberish suffixes that jailbreak aligned models."},"explain":{"zh":"用梯度引导的贪心坐标搜索优化提示后缀，使模型以“当然，以下是…”开头作答。找到的后缀可迁移到闭源模型，揭示对齐的脆弱性。","en":"Greedy coordinate gradient search optimizes a suffix so the model starts with an affirmative reply; suffixes transfer to closed models, exposing fragile alignment."},"paper":{"title":"Universal and Transferable Adversarial Attacks on Aligned Language Models","authors":["Andy Zou","Zifan Wang","Nicholas Carlini","et al."],"firstPublished":"2023-07-27","year":2023,"arxiv":"2307.15043","url":"https://arxiv.org/abs/2307.15043","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["adversarial-examples","rlhf"],"usedBy":[],"related":["red-teaming","prompt-injection"],"usedIn":{"zh":"模型安全评测与防御研究","en":"Safety evaluation and defense research"},"resources":[{"title":"llm-attacks（GCG 官方实现）","url":"https://github.com/llm-attacks/llm-attacks","type":"repo","lang":"en"},{"title":"Lilian Weng：Adversarial Attacks on LLMs","url":"https://lilianweng.github.io/posts/2023-10-25-adv-attack-llm/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/gcg-jailbreak/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/gcg-jailbreak/"}},{"id":"prompt-injection","name":{"zh":"间接提示注入","en":"Indirect Prompt Injection"},"aka":[],"category":"safety","tier":"core","year":2023,"difficulty":2,"summary":{"zh":"把恶意指令藏进网页、邮件或文档，借模型之手执行。","en":"Hides malicious instructions in content the model reads."},"explain":{"zh":"当 Agent 读取外部数据时，数据里的“指令”可能被当成用户命令执行，导致泄露数据或执行越权操作。它是 Agent 与 RAG 应用的头号安全风险。","en":"When agents read external data, embedded instructions may be followed as commands, causing leaks or unauthorized actions — the top risk for agents and RAG apps."},"paper":{"title":"Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection","authors":["Kai Greshake","Sahar Abdelnabi","Shailesh Mishra","et al."],"firstPublished":"2023-02-23","year":2023,"arxiv":"2302.12173","url":"https://arxiv.org/abs/2302.12173","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["rag","function-calling"],"usedBy":[],"related":["gcg-jailbreak","red-teaming"],"usedIn":{"zh":"Agent、浏览器助手、邮件助手的安全设计","en":"Security of agents, browser and email assistants"},"resources":[{"title":"OWASP Top 10 for LLM Applications","url":"https://genai.owasp.org/llm-top-10/","type":"docs","lang":"en"},{"title":"Lilian Weng：Adversarial Attacks on LLMs","url":"https://lilianweng.github.io/posts/2023-10-25-adv-attack-llm/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/prompt-injection/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/prompt-injection/"}},{"id":"red-teaming","name":{"zh":"自动化红队","en":"Automated Red Teaming"},"aka":[],"category":"safety","tier":"core","year":2022,"difficulty":2,"summary":{"zh":"用一个语言模型去生成攻击提示，测试另一个模型。","en":"Uses one LM to generate test cases that attack another."},"explain":{"zh":"红队模型自动生成大量可能诱发有害、泄露隐私或冒犯性回答的问题，由分类器判断是否成功，比人工红队覆盖更广。","en":"A red-team LM generates many prompts likely to elicit harmful or private outputs; a classifier scores success, covering far more than manual testing."},"paper":{"title":"Red Teaming Language Models with Language Models","authors":["Ethan Perez","Saffron Huang","Francis Song","et al."],"firstPublished":"2022-02-07","year":2022,"arxiv":"2202.03286","url":"https://arxiv.org/abs/2202.03286","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["rlhf"],"usedBy":[],"related":["gcg-jailbreak","llm-judge"],"usedIn":{"zh":"前沿模型发布前安全评估","en":"Pre-deployment safety evaluation"},"resources":[{"title":"Lilian Weng：Adversarial Attacks on LLMs","url":"https://lilianweng.github.io/posts/2023-10-25-adv-attack-llm/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/red-teaming/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/red-teaming/"}},{"id":"llm-judge","name":{"zh":"大模型当评委","en":"LLM-as-a-Judge"},"aka":[],"category":"safety","tier":"core","year":2023,"difficulty":2,"summary":{"zh":"用强模型给其他模型的回答打分或做两两比较。","en":"Strong LLMs grade or compare other models' answers."},"explain":{"zh":"在 MT-Bench 和 Chatbot Arena 上验证，GPT-4 评委与人类偏好一致率超过 80%，但存在位置偏差、长度偏差和自我偏好，需要交换顺序等校正。","en":"On MT-Bench and Chatbot Arena, GPT-4 judges agree with humans over 80%, but show position, verbosity and self-preference biases that need mitigation."},"paper":{"title":"Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena","authors":["Lianmin Zheng","Wei-Lin Chiang","Ying Sheng","et al."],"firstPublished":"2023-06-09","year":2023,"arxiv":"2306.05685","url":"https://arxiv.org/abs/2306.05685","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["rlhf"],"usedBy":[],"related":["red-teaming","reward-model"],"usedIn":{"zh":"模型评测、RAG/Agent 自动评估","en":"Model evaluation, automatic RAG/agent evals"},"resources":[{"title":"lm-sys/FastChat（MT-Bench）","url":"https://github.com/lm-sys/FastChat","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/llm-judge/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/llm-judge/"}},{"id":"watermarking","name":{"zh":"大模型文本水印","en":"LLM Watermarking"},"aka":[],"category":"safety","tier":"core","year":2023,"difficulty":3,"summary":{"zh":"生成时暗中偏向一组“绿色”词，事后可统计检测 AI 文本。","en":"Biases generation toward a hidden green list so AI text is detectable."},"explain":{"zh":"用前一个 token 做种子把词表随机分成绿/红两组，生成时略微提高绿词概率；检测时统计绿词比例，人类几乎察觉不到但可高置信度识别。","en":"Seed a random green/red vocabulary split from the previous token and slightly boost green tokens; detection counts green tokens with high confidence."},"paper":{"title":"A Watermark for Large Language Models","authors":["John Kirchenbauer","Jonas Geiping","Yuxin Wen","et al."],"firstPublished":"2023-01-24","year":2023,"arxiv":"2301.10226","url":"https://arxiv.org/abs/2301.10226","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["sampling"],"usedBy":[],"related":["red-teaming"],"usedIn":{"zh":"AI 内容溯源（如 Google SynthID）","en":"AI content provenance (e.g., Google SynthID)"},"resources":[{"title":"jwkirchenbauer/lm-watermarking（官方实现）","url":"https://github.com/jwkirchenbauer/lm-watermarking","type":"repo","lang":"en"},{"title":"Google DeepMind：SynthID","url":"https://deepmind.google/models/synthid/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/watermarking/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/watermarking/"}},{"id":"dp-sgd","name":{"zh":"差分隐私 SGD","en":"Differentially Private SGD (DP-SGD)"},"aka":[],"category":"safety","tier":"extended","year":2016,"difficulty":4,"summary":{"zh":"训练时裁剪每个样本的梯度并加噪声，保证不泄露个体数据。","en":"Clips per-example gradients and adds noise to protect individuals."},"explain":{"zh":"每个样本的梯度先裁剪到固定范数，再在汇总时加入高斯噪声，并用“矩会计”跟踪总隐私预算，使模型难以记住任何单个训练样本。","en":"Clip each example's gradient, add Gaussian noise to the sum, and track privacy budget with the moments accountant, so no single example is memorized."},"paper":{"title":"Deep Learning with Differential Privacy","authors":["Martín Abadi","Andy Chu","Ian Goodfellow","et al."],"firstPublished":"2016-07-01","year":2016,"arxiv":"1607.00133","url":"https://arxiv.org/abs/1607.00133","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["sgd"],"usedBy":[],"related":["watermarking","membership-inference"],"usedIn":{"zh":"医疗、金融、键盘输入法等隐私数据训练","en":"Training on medical, financial or keyboard data"},"resources":[{"title":"Opacus（PyTorch 差分隐私库）","url":"https://opacus.ai/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/dp-sgd/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/dp-sgd/"}},{"id":"weak-to-strong","name":{"zh":"弱到强泛化","en":"Weak-to-Strong Generalization"},"aka":[],"category":"safety","tier":"core","year":2023,"difficulty":4,"summary":{"zh":"用弱模型的标签监督强模型，研究人类如何监督超人 AI。","en":"Supervising strong models with weak-model labels as an analogy for superalignment."},"explain":{"zh":"让 GPT-2 级别模型给 GPT-4 级别模型提供标签，发现强模型能超越弱老师，但仍未发挥全部能力；加入辅助置信度损失可缩小差距。","en":"GPT-2-level labels supervise a GPT-4-level student; the student beats its teacher but not its full potential, and an auxiliary confidence loss narrows the gap."},"paper":{"title":"Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision","authors":["Collin Burns","Pavel Izmailov","Jan Hendrik Kirchner","et al."],"firstPublished":"2023-12-14","year":2023,"arxiv":"2312.09390","url":"https://arxiv.org/abs/2312.09390","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["rlhf","knowledge-distillation"],"usedBy":[],"related":["debate"],"usedIn":{"zh":"可扩展监督研究","en":"Scalable oversight research"},"resources":[{"title":"OpenAI：Weak-to-strong generalization","url":"https://openai.com/index/weak-to-strong-generalization/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/weak-to-strong/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/weak-to-strong/"}},{"id":"debate","name":{"zh":"AI 辩论","en":"AI Safety via Debate"},"aka":[],"category":"safety","tier":"extended","year":2018,"difficulty":3,"summary":{"zh":"让两个 AI 辩论，由人类裁判，帮助人类监督复杂问题。","en":"Two AIs debate and a human judges, aiding oversight of hard questions."},"explain":{"zh":"即便人类无法直接判断复杂答案的对错，也可以通过观察两个 AI 互相指出对方漏洞来做出更准确的判断。它是可扩展监督的代表方案之一。","en":"Even when humans can't verify an answer directly, judging two AIs exposing each other's flaws can be more reliable — a key scalable-oversight proposal."},"paper":{"title":"AI safety via debate","authors":["Geoffrey Irving","Paul Christiano","Dario Amodei"],"firstPublished":"2018-05-02","year":2018,"arxiv":"1805.00899","url":"https://arxiv.org/abs/1805.00899","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["rlhf"],"usedBy":[],"related":["weak-to-strong","constitutional-ai"],"usedIn":{"zh":"可扩展监督研究","en":"Scalable oversight research"},"resources":[{"title":"Nathan Lambert：RLHF Book","url":"https://rlhfbook.com/","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/debate/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/debate/"}},{"id":"sleeper-agents","name":{"zh":"潜伏后门与欺骗性对齐","en":"Sleeper Agents (Backdoors that Persist)"},"aka":[],"category":"safety","tier":"extended","year":2024,"difficulty":4,"summary":{"zh":"植入后门的模型能在安全训练后依然保持隐藏的恶意行为。","en":"Backdoored models keep hidden behaviors even after safety training."},"explain":{"zh":"作者训练出在“年份为 2024”时写入漏洞代码的模型，发现 SFT、RLHF 和对抗训练都难以消除，甚至让模型更会隐藏触发条件。提醒我们安全训练可能制造“安全的假象”。","en":"Models trained to insert vulnerable code when the year is 2024 kept this through SFT, RLHF and adversarial training, which could even teach better concealment — warning of a false sense of safety."},"paper":{"title":"Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training","authors":["Evan Hubinger","Carson Denison","Jesse Mu","et al."],"firstPublished":"2024-01-10","year":2024,"arxiv":"2401.05566","url":"https://arxiv.org/abs/2401.05566","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-21"},"prereqs":["rlhf","adversarial-examples"],"usedBy":[],"related":["interp-sae","weak-to-strong"],"usedIn":{"zh":"前沿 AI 安全评估","en":"Frontier AI safety evaluation"},"resources":[{"title":"Transformer Circuits Thread","url":"https://transformer-circuits.pub/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/sleeper-agents/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/sleeper-agents/"}},{"id":"looped-lm","name":{"zh":"循环语言模型（Looped LM）","en":"Looped Language Models"},"aka":["Universal Transformer","Recurrent-depth Transformer","循环深度"],"category":"new-arch","tier":"extended","year":2018,"difficulty":4,"summary":{"zh":"同一组层反复执行多次，用“算更多遍”代替“堆更多层”。","en":"Re-applies the same layers several times, trading depth in parameters for depth in compute."},"explain":{"zh":"直觉：与其堆 100 层不同参数，不如让同一个小模块把表示反复“过几遍”，每过一遍就多想一步。这样参数量不变，却能在推理时按需多循环以换取更强推理。Universal Transformer 最早提出这一做法，2025 年的 Ouro 把它扩展到预训练规模；代价是每多循环一次，训练和解码成本都增加。","en":"Intuition: instead of stacking 100 distinct layers, run one block over the representation several times, each pass adding a step of thought. Parameters stay fixed while extra loops buy more reasoning at inference. Universal Transformers introduced it; Ouro (2025) scaled it to pre-training; every extra loop costs training and decoding compute."},"paper":{"title":"Universal Transformers","authors":["Mostafa Dehghani","Stephan Gouws","Oriol Vinyals","et al."],"firstPublished":"2018-07-10","year":2018,"arxiv":"1807.03819","url":"https://arxiv.org/abs/1807.03819","source":"DataCite (arXiv DOI 10.48550/arXiv.1807.03819)","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["transformer"],"usedBy":[],"related":["mamba","mtp"],"usedIn":{"zh":"Ouro 循环语言模型（字节跳动 Seed 等）；2026-10 的 LEV 循环决策模型、不动点训练研究，以及 PAR 进度对齐循环（2610.27628）","en":"Ouro looped LMs (ByteDance Seed et al.); Oct 2026 work on looped decision models (LEV), fixed-point training and progress-aligned recurrence (PAR, 2610.27628)"},"resources":[{"title":"Ouro 循环语言模型项目页","url":"https://ouro-llm.github.io/","type":"docs","lang":"en"},{"title":"ByteDance/Ouro-1.4B 模型卡","url":"https://huggingface.co/ByteDance/Ouro-1.4B","type":"repo","lang":"en"},{"title":"tensorflow/tensor2tensor（含 Universal Transformer 原始实现，已归档）","url":"https://github.com/tensorflow/tensor2tensor","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/looped-lm/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/looped-lm/"}},{"id":"lasso","name":{"zh":"正则化回归（Lasso / Ridge）","en":"Regularized Regression (Lasso / Ridge)"},"aka":["Ridge","L1/L2 Regularization","Elastic Net","岭回归"],"category":"classic-ml","tier":"core","year":1996,"difficulty":2,"summary":{"zh":"给线性模型的权重加惩罚，防止过拟合并自动筛选特征。","en":"Penalize linear-model weights to curb overfitting and select features."},"explain":{"zh":"在“误差尽量小”之外再要求“权重尽量小”。L2 惩罚（Ridge）让权重整体变小，L1 惩罚（Lasso）会把不重要特征的权重压到 0，相当于自动做特征选择。惩罚强度是要调的超参数，这也是理解“正则化”最直观的入口。","en":"Besides small error, also ask for small weights. An L2 penalty (Ridge) shrinks all weights; an L1 penalty (Lasso) drives unimportant ones to exactly zero, selecting features automatically. The penalty strength is a hyperparameter – the clearest entry point to regularization."},"paper":{"title":"Regression Shrinkage and Selection Via the Lasso","authors":["Robert Tibshirani"],"venue":"Journal of the Royal Statistical Society Series B","year":1996,"doi":"10.1111/j.2517-6161.1996.tb02080.x","url":"https://doi.org/10.1111/j.2517-6161.1996.tb02080.x","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["linear-regression"],"usedBy":[],"related":["logistic-regression","svm"],"usedIn":{"zh":"高维数据的特征选择、稀疏回归、基因表达分析","en":"Feature selection on high-dimensional data, sparse regression, gene-expression analysis"},"resources":[{"title":"scikit-learn 用户指南：线性模型","url":"https://scikit-learn.org/stable/modules/linear_model.html","type":"docs","lang":"en"},{"title":"An Introduction to Statistical Learning（ISLP，免费）","url":"https://www.statlearning.com/","type":"book","lang":"en"},{"title":"StatQuest 视频索引","url":"https://statquest.org/video-index/","type":"video","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/lasso/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/lasso/"}},{"id":"adaboost","name":{"zh":"AdaBoost","en":"AdaBoost"},"aka":[],"category":"classic-ml","tier":"core","year":1997,"difficulty":2,"summary":{"zh":"串联许多弱分类器，每轮更关注上一轮分错的样本。","en":"Chains weak learners, each focusing on the examples the previous ones got wrong."},"explain":{"zh":"先训练一个很弱的分类器，把它分错的样本权重调高，再训练下一个，如此反复，最后按各自准确率加权投票。它是“提升（Boosting）”思想的第一个实用算法，后来的梯度提升树（GBDT）把它推广到任意损失函数。","en":"Train a weak classifier, up-weight the examples it misclassifies, train the next one, repeat, then take an accuracy-weighted vote. It was the first practical boosting algorithm; gradient boosting (GBDT) later generalized it to any loss."},"paper":{"title":"A Decision-Theoretic Generalization of On-Line Learning and an Application to Boosting","authors":["Yoav Freund","Robert E. Schapire"],"venue":"Journal of Computer and System Sciences","year":1997,"doi":"10.1006/jcss.1997.1504","url":"https://doi.org/10.1006/jcss.1997.1504","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["decision-tree"],"usedBy":[],"related":["gbdt","random-forest"],"usedIn":{"zh":"Viola–Jones 人脸检测（2001）、表格数据分类基线","en":"Viola–Jones face detection (2001), tabular classification baselines"},"resources":[{"title":"scikit-learn 用户指南：集成方法","url":"https://scikit-learn.org/stable/modules/ensemble.html","type":"docs","lang":"en"},{"title":"StatQuest 视频索引","url":"https://statquest.org/video-index/","type":"video","lang":"en"},{"title":"The Elements of Statistical Learning（免费）","url":"https://hastie.su.domains/ElemStatLearn/","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/adaboost/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/adaboost/"}},{"id":"dbscan","name":{"zh":"DBSCAN 密度聚类","en":"DBSCAN (Density-Based Clustering)"},"aka":[],"category":"classic-ml","tier":"extended","year":1996,"difficulty":2,"summary":{"zh":"按“点够不够密”来聚类，能找出任意形状的簇和噪声点。","en":"Clusters by local density; finds arbitrarily shaped clusters and flags noise."},"explain":{"zh":"一个点周围半径 ε 内的邻居足够多，就算“核心点”；彼此可达的核心点连成一个簇，孤立的点标为噪声。它不需要事先指定簇的个数，能处理 k-means 搞不定的非球形簇，但对 ε 和最少邻居数比较敏感。","en":"A point with enough neighbours within radius ε is a core point; core points that reach each other form a cluster and isolated points are noise. No need to choose the number of clusters, and it handles non-spherical shapes that k-means cannot, but it is sensitive to ε and the minimum-neighbour count."},"paper":{"title":"A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise","authors":["Martin Ester","Hans-Peter Kriegel","Jörg Sander","Xiaowei Xu"],"venue":"KDD-96","year":1996,"url":"https://cdn.aaai.org/KDD/1996/KDD96-037.pdf","source":"官方页面 HTTP 核实","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["kmeans"],"usedBy":[],"related":["umap","em"],"usedIn":{"zh":"地理位置聚类、异常检测、嵌入向量的聚类分析","en":"Geospatial clustering, anomaly detection, clustering embedding vectors"},"resources":[{"title":"scikit-learn 用户指南：DBSCAN","url":"https://scikit-learn.org/stable/modules/clustering.html#dbscan","type":"docs","lang":"en"},{"title":"StatQuest 视频索引","url":"https://statquest.org/video-index/","type":"video","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/dbscan/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/dbscan/"}},{"id":"naive-bayes","name":{"zh":"朴素贝叶斯","en":"Naive Bayes"},"aka":[],"category":"classic-ml","tier":"core","year":1997,"difficulty":1,"summary":{"zh":"假设特征相互独立，用贝叶斯公式算出最可能的类别。","en":"Applies Bayes’ rule assuming features are independent given the class."},"explain":{"zh":"用训练数据统计“每个类别下每个特征出现的概率”，预测时把这些概率连乘再乘上类别先验，取最大的类别。“特征相互独立”的假设几乎从不成立，但它训练极快、数据少时也稳，长期是文本分类的强基线。","en":"Count how likely each feature is under each class, then at prediction time multiply those probabilities with the class prior and pick the largest. The independence assumption is almost never true, yet it trains instantly, works with little data, and has long been a strong text-classification baseline."},"paper":{"title":"On the Optimality of the Simple Bayesian Classifier under Zero-One Loss","authors":["Pedro Domingos","Michael Pazzani"],"venue":"Machine Learning","year":1997,"doi":"10.1023/A:1007413511361","url":"https://doi.org/10.1023/A:1007413511361","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":[],"usedBy":[],"related":["logistic-regression","knn"],"usedIn":{"zh":"垃圾邮件过滤、文本分类基线","en":"Spam filtering, text-classification baselines"},"resources":[{"title":"scikit-learn 用户指南：朴素贝叶斯","url":"https://scikit-learn.org/stable/modules/naive_bayes.html","type":"docs","lang":"en"},{"title":"StatQuest 视频索引","url":"https://statquest.org/video-index/","type":"video","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/naive-bayes/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/naive-bayes/"}},{"id":"gru","name":{"zh":"门控循环单元（GRU）","en":"Gated Recurrent Unit (GRU)"},"aka":[],"category":"dl-foundations","tier":"core","year":2014,"difficulty":2,"summary":{"zh":"比 LSTM 更简洁的门控 RNN，用两个门控制记住和遗忘。","en":"A leaner gated RNN than LSTM, with two gates for remembering and forgetting."},"explain":{"zh":"GRU 把 LSTM 的三个门合并成“更新门”和“重置门”，也不再单独保存细胞状态。参数更少、训练更快，在很多序列任务上效果和 LSTM 相当。它和编码器-解码器结构出自同一篇论文。","en":"GRU merges LSTM’s three gates into an update gate and a reset gate and drops the separate cell state. Fewer parameters, faster training, and on many sequence tasks it matches LSTM. It was introduced in the same paper as the RNN encoder–decoder."},"paper":{"title":"Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation","authors":["Kyunghyun Cho","Bart van Merrienboer","Caglar Gulcehre","et al."],"firstPublished":"2014-06-03","year":2014,"arxiv":"1406.1078","url":"https://arxiv.org/abs/1406.1078","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["lstm"],"usedBy":[],"related":["seq2seq","bahdanau-attention"],"usedIn":{"zh":"轻量序列建模、时间序列预测、早期神经机器翻译","en":"Lightweight sequence modelling, time-series forecasting, early neural machine translation"},"resources":[{"title":"D2L：门控循环单元 GRU","url":"https://d2l.ai/chapter_recurrent-modern/gru.html","type":"book","lang":"en"},{"title":"PyTorch 文档：nn.GRU","url":"https://docs.pytorch.org/docs/stable/generated/torch.nn.GRU.html","type":"docs","lang":"en"},{"title":"colah：Understanding LSTM Networks","url":"https://colah.github.io/posts/2015-08-Understanding-LSTMs/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/gru/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/gru/"}},{"id":"autoencoder","name":{"zh":"自编码器","en":"Autoencoder"},"aka":[],"category":"dl-foundations","tier":"core","year":2006,"difficulty":2,"summary":{"zh":"先把输入压缩成短向量再还原，学到数据的紧凑表示。","en":"Compress the input to a short code and reconstruct it, learning a compact representation."},"explain":{"zh":"网络分两半：编码器把输入压成低维向量，解码器再把它还原。训练目标就是“还原得越像越好”，中间那个瓶颈迫使网络只保留最重要的信息。它可以看作非线性的 PCA，也是 VAE、MAE 和 Stable Diffusion 潜空间的源头。","en":"An encoder squeezes the input into a low-dimensional code and a decoder reconstructs it; training just minimizes reconstruction error, so the bottleneck keeps only what matters. Think of it as nonlinear PCA – and the ancestor of VAEs, MAE and the latent space of Stable Diffusion."},"paper":{"title":"Reducing the Dimensionality of Data with Neural Networks","authors":["G. E. Hinton","R. R. Salakhutdinov"],"venue":"Science","year":2006,"doi":"10.1126/science.1127647","url":"https://doi.org/10.1126/science.1127647","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["backprop"],"usedBy":[],"related":["pca","vae","mae"],"usedIn":{"zh":"降维、异常检测、图像去噪；潜扩散模型的图像编码器","en":"Dimensionality reduction, anomaly detection, denoising; the image encoder of latent diffusion models"},"resources":[{"title":"Goodfellow 等：Deep Learning","url":"https://www.deeplearningbook.org/","type":"book","lang":"en"},{"title":"Lilian Weng：From Autoencoder to Beta-VAE","url":"https://lilianweng.github.io/posts/2018-08-12-vae/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/autoencoder/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/autoencoder/"}},{"id":"unet","name":{"zh":"U-Net","en":"U-Net"},"aka":[],"category":"dl-foundations","tier":"core","year":2015,"difficulty":2,"summary":{"zh":"先下采样再上采样、中间用跳连拼接的“U 形”卷积网络。","en":"A U-shaped conv net: downsample, upsample, and skip-connect matching levels."},"explain":{"zh":"左半边像普通 CNN 一样逐层缩小、提取语义，右半边再逐层放大回原尺寸；同一层级的左右特征直接拼接，让输出既懂语义又保留细节。它原本用于医学图像分割，后来成为 DDPM 和 Stable Diffusion 去噪网络的骨架。","en":"The left half shrinks the image like a normal CNN to capture meaning; the right half upsamples back to full size, and skip connections concatenate same-level features so the output keeps both semantics and detail. Built for biomedical segmentation, it later became the denoiser backbone of DDPM and Stable Diffusion."},"paper":{"title":"U-Net: Convolutional Networks for Biomedical Image Segmentation","authors":["Olaf Ronneberger","Philipp Fischer","Thomas Brox"],"firstPublished":"2015-05-18","year":2015,"arxiv":"1505.04597","url":"https://arxiv.org/abs/1505.04597","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["cnn"],"usedBy":[],"related":["ddpm","latent-diffusion","sam"],"usedIn":{"zh":"医学图像分割；DDPM、Stable Diffusion 1.x/2.x 的去噪网络","en":"Medical image segmentation; the denoising network in DDPM and Stable Diffusion 1.x/2.x"},"resources":[{"title":"Diffusers 文档","url":"https://huggingface.co/docs/diffusers/index","type":"docs","lang":"en"},{"title":"CompVis/latent-diffusion（官方实现）","url":"https://github.com/CompVis/latent-diffusion","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/unet/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/unet/"}},{"id":"weight-init","name":{"zh":"权重初始化（Xavier / He）","en":"Weight Initialization (Xavier / He)"},"aka":["Kaiming Init","Glorot Init","He 初始化"],"category":"dl-foundations","tier":"core","year":2010,"difficulty":2,"summary":{"zh":"按层宽设定初始权重的方差，让信号和梯度不爆炸也不消失。","en":"Scale initial weights by layer width so signals and gradients neither explode nor vanish."},"explain":{"zh":"如果初始权重太大，信号逐层放大；太小，又逐层衰减到 0，深网络就训不动。Xavier 初始化按输入、输出宽度设定方差，适合 tanh；He 初始化把方差再乘 2，适合 ReLU。现在各框架的默认初始化都来自这两种思路。","en":"Weights that start too large blow signals up layer after layer; too small and they fade to zero, so deep nets will not train. Xavier init sets the variance from fan-in and fan-out (good for tanh); He init doubles it for ReLU. Framework defaults today descend from these two ideas."},"paper":{"title":"Understanding the difficulty of training deep feedforward neural networks","authors":["Xavier Glorot","Yoshua Bengio"],"venue":"AISTATS 2010 (PMLR v9)","year":2010,"url":"https://proceedings.mlr.press/v9/glorot10a.html","source":"官方页面 HTTP 核实","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["backprop"],"usedBy":[],"related":["batchnorm","resnet","mup"],"usedIn":{"zh":"PyTorch、JAX 等框架层的默认初始化","en":"Default layer initializers in PyTorch, JAX and other frameworks"},"resources":[{"title":"PyTorch 文档：torch.nn.init","url":"https://docs.pytorch.org/docs/stable/nn.init.html","type":"docs","lang":"en"},{"title":"D2L：数值稳定性与模型初始化","url":"https://d2l.ai/chapter_multilayer-perceptrons/numerical-stability-and-init.html","type":"book","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/weight-init/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/weight-init/"}},{"id":"glove","name":{"zh":"GloVe 词向量","en":"GloVe"},"aka":[],"category":"dl-foundations","tier":"extended","year":2014,"difficulty":1,"summary":{"zh":"用全局词共现统计训练词向量，和 Word2Vec 齐名。","en":"Word vectors trained on global word co-occurrence counts; Word2Vec’s main peer."},"explain":{"zh":"先统计整个语料中“两个词一起出现多少次”，再训练词向量，使两个向量的点积接近共现次数的对数。它把计数法和预测法的优点结合起来，训练出的向量同样能做“国王 − 男人 + 女人 ≈ 女王”这类类比。","en":"First count how often every pair of words co-occurs in the corpus, then fit vectors whose dot product approximates the log count. It blends count-based and prediction-based methods, and its vectors support the same “king − man + woman ≈ queen” analogies."},"paper":{"title":"GloVe: Global Vectors for Word Representation","authors":["Jeffrey Pennington","Richard Socher","Christopher Manning"],"venue":"EMNLP 2014","year":2014,"doi":"10.3115/v1/D14-1162","url":"https://doi.org/10.3115/v1/D14-1162","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["word2vec"],"usedBy":[],"related":["elmo"],"usedIn":{"zh":"预训练词向量基线、NLP 教学","en":"Pre-trained word-vector baselines, NLP teaching"},"resources":[{"title":"Stanford NLP：GloVe 项目主页","url":"https://nlp.stanford.edu/projects/glove/","type":"docs","lang":"en"},{"title":"Stanford CS224N 深度学习自然语言处理","url":"https://web.stanford.edu/class/cs224n/","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/glove/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/glove/"}},{"id":"elmo","name":{"zh":"ELMo 上下文词向量","en":"ELMo (Contextual Word Embeddings)"},"aka":[],"category":"dl-foundations","tier":"extended","year":2018,"difficulty":2,"summary":{"zh":"用双向 LSTM 语言模型，让同一个词在不同句子里有不同向量。","en":"Bidirectional LSTM language-model features: the same word gets different vectors in different contexts."},"explain":{"zh":"Word2Vec 给每个词一个固定向量，“苹果”在水果和手机语境里一样。ELMo 先训练一个双向 LSTM 语言模型，再把各层隐状态加权组合作为词的表示，于是词义随上下文变化。它证明了“先预训练语言模型、再迁移”的威力，几个月后被 BERT 接棒。","en":"Word2Vec gives each word one fixed vector, whatever the context. ELMo pre-trains a bidirectional LSTM language model and uses a learned mix of its layers as the word’s representation, so meaning shifts with context. It proved the power of pre-train-then-transfer, and BERT took over months later."},"paper":{"title":"Deep contextualized word representations","authors":["Matthew E. Peters","Mark Neumann","Mohit Iyyer","et al."],"firstPublished":"2018-02-15","year":2018,"arxiv":"1802.05365","url":"https://arxiv.org/abs/1802.05365","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["lstm","word2vec"],"usedBy":[],"related":["bert","glove"],"usedIn":{"zh":"2018 年多项 NLP 基准的上下文特征；BERT 之前的代表方法","en":"Contextual features behind several 2018 NLP benchmark gains; the key pre-BERT method"},"resources":[{"title":"Stanford CS224N 深度学习自然语言处理","url":"https://web.stanford.edu/class/cs224n/","type":"course","lang":"en"},{"title":"Jay Alammar：The Illustrated BERT","url":"https://jalammar.github.io/illustrated-bert/","type":"blog","lang":"en"},{"title":"allenai/allennlp-models（ELMo 参考实现，已归档）","url":"https://github.com/allenai/allennlp-models","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/elmo/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/elmo/"}},{"id":"sentencepiece","name":{"zh":"SentencePiece 分词","en":"SentencePiece Tokenization"},"aka":["Unigram LM Tokenizer"],"category":"transformer","tier":"core","year":2018,"difficulty":2,"summary":{"zh":"直接在原始文本上训练子词分词器，不依赖空格，适合中文等语言。","en":"Trains subword tokenizers on raw text without relying on spaces – works for Chinese and other languages."},"explain":{"zh":"它把空格也当成普通字符，直接从原始句子学习子词表，支持 BPE 和 Unigram 两种算法，同一套规则既能切分也能无损还原。不用先做按语言定制的预分词，因此对中文、日文这类没有空格的语言特别友好。","en":"It treats the space as an ordinary symbol and learns a subword vocabulary straight from raw sentences, supporting both BPE and Unigram, with lossless detokenization. No language-specific pre-tokenization is needed, which suits languages without spaces such as Chinese and Japanese."},"paper":{"title":"SentencePiece: A simple and language independent subword tokenizer and detokenizer for Neural Text Processing","authors":["Taku Kudo","John Richardson"],"firstPublished":"2018-08-19","year":2018,"arxiv":"1808.06226","url":"https://arxiv.org/abs/1808.06226","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["bpe"],"usedBy":[],"related":["t5"],"usedIn":{"zh":"T5、Llama 1/2、Gemma 等模型的分词器","en":"The tokenizer used by T5, Llama 1/2, Gemma and others"},"resources":[{"title":"google/sentencepiece","url":"https://github.com/google/sentencepiece","type":"repo","lang":"en"},{"title":"HF LLM 课程：BPE 分词","url":"https://huggingface.co/learn/llm-course/chapter6/5","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/sentencepiece/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/sentencepiece/"}},{"id":"faster-rcnn","name":{"zh":"Faster R-CNN 目标检测","en":"Faster R-CNN (Object Detection)"},"aka":["R-CNN","Mask R-CNN","RPN"],"category":"multimodal","tier":"core","year":2015,"difficulty":3,"summary":{"zh":"先由网络提出候选框，再分类并精修位置的两阶段检测器。","en":"Two-stage detector: a network proposes boxes, then classifies and refines them."},"explain":{"zh":"第一阶段的区域建议网络（RPN）在特征图上滑动，给出可能有物体的候选框；第二阶段对每个候选框判断类别、微调边框。它把候选框生成也交给神经网络，速度和精度都大幅提升，是两阶段检测器的代表，Mask R-CNN 在它的基础上加了实例分割。","en":"Stage one, a Region Proposal Network, slides over the feature map and proposes boxes likely to contain objects; stage two classifies each box and refines it. Handing proposals to a neural network made detection much faster and more accurate; it is the canonical two-stage detector, and Mask R-CNN builds on it for instance segmentation."},"paper":{"title":"Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks","authors":["Shaoqing Ren","Kaiming He","Ross Girshick","et al."],"firstPublished":"2015-06-04","year":2015,"arxiv":"1506.01497","url":"https://arxiv.org/abs/1506.01497","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["cnn","resnet"],"usedBy":["detr"],"related":["yolo","detr","sam"],"usedIn":{"zh":"Detectron2、torchvision 中的标准检测器；工业质检、自动驾驶感知","en":"A standard detector in Detectron2 and torchvision; industrial inspection, driving perception"},"resources":[{"title":"torchvision 文档：目标检测与实例分割模型","url":"https://docs.pytorch.org/vision/stable/models.html#object-detection-instance-segmentation-and-person-keypoint-detection","type":"docs","lang":"en"},{"title":"facebookresearch/detectron2","url":"https://github.com/facebookresearch/detectron2","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/faster-rcnn/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/faster-rcnn/"}},{"id":"yolo","name":{"zh":"YOLO 实时目标检测","en":"YOLO (Real-Time Object Detection)"},"aka":[],"category":"multimodal","tier":"core","year":2015,"difficulty":2,"summary":{"zh":"一次前向就同时预测所有物体的框和类别，速度极快。","en":"Predicts all boxes and classes in a single forward pass – very fast."},"explain":{"zh":"把图像划成网格，每个格子直接回归若干个边框和类别概率，不需要单独的候选框阶段。准确率起初不如两阶段方法，但速度快一个数量级，可以实时跑视频。此后十年的 YOLO 系列不断改进，成了工业界最常用的检测器家族。","en":"Split the image into a grid and have each cell directly regress boxes and class probabilities – no separate proposal stage. Early versions trailed two-stage detectors in accuracy but ran an order of magnitude faster, enough for real-time video. A decade of YOLO successors made it the most widely used detector family in industry."},"paper":{"title":"You Only Look Once: Unified, Real-Time Object Detection","authors":["Joseph Redmon","Santosh Divvala","Ross Girshick","et al."],"firstPublished":"2015-06-08","year":2015,"arxiv":"1506.02640","url":"https://arxiv.org/abs/1506.02640","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["cnn"],"usedBy":[],"related":["faster-rcnn","detr"],"usedIn":{"zh":"视频监控、无人机、工业质检等实时检测（Ultralytics YOLO 系列）","en":"Real-time detection in video analytics, drones and inspection (Ultralytics YOLO family)"},"resources":[{"title":"Ultralytics YOLO 官方文档","url":"https://docs.ultralytics.com/","type":"docs","lang":"multi"},{"title":"YOLO 作者项目页（Darknet）","url":"https://pjreddie.com/darknet/yolo/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/yolo/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/yolo/"}},{"id":"detr","name":{"zh":"DETR（Transformer 目标检测）","en":"DETR (Detection Transformer)"},"aka":[],"category":"multimodal","tier":"core","year":2020,"difficulty":3,"summary":{"zh":"用 Transformer 把检测变成“集合预测”，去掉锚框和 NMS。","en":"Casts detection as set prediction with a Transformer – no anchors, no NMS."},"explain":{"zh":"CNN 提特征后交给 Transformer，固定数量的“物体查询”各自输出一个框或“无物体”，训练时用二分图匹配把预测和真实框一一对应。它不再需要手工设计的锚框和非极大值抑制，结构简洁，开启了 Transformer 检测器这条路线。","en":"A CNN extracts features, a Transformer turns a fixed set of object queries into boxes (or “no object”), and bipartite matching pairs predictions with ground truth during training. Hand-crafted anchors and non-maximum suppression disappear, opening the line of Transformer-based detectors."},"paper":{"title":"End-to-End Object Detection with Transformers","authors":["Nicolas Carion","Francisco Massa","Gabriel Synnaeve","et al."],"firstPublished":"2020-05-26","year":2020,"arxiv":"2005.12872","url":"https://arxiv.org/abs/2005.12872","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["transformer","faster-rcnn"],"usedBy":[],"related":["vit","sam"],"usedIn":{"zh":"Deformable DETR、RT-DETR、开放词表检测等后续工作的基础","en":"The basis for Deformable DETR, RT-DETR and open-vocabulary detectors"},"resources":[{"title":"facebookresearch/detr（官方实现）","url":"https://github.com/facebookresearch/detr","type":"repo","lang":"en"},{"title":"Hugging Face Transformers：DETR","url":"https://huggingface.co/docs/transformers/model_doc/detr","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/detr/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/detr/"}},{"id":"swin","name":{"zh":"Swin Transformer","en":"Swin Transformer"},"aka":[],"category":"multimodal","tier":"extended","year":2021,"difficulty":3,"summary":{"zh":"在局部窗口里算注意力并逐层合并，适合检测分割的层级视觉骨干。","en":"Hierarchical vision Transformer with shifted local-window attention."},"explain":{"zh":"ViT 对整张图做全局注意力，分辨率一高计算量就爆炸。Swin 只在小窗口内算注意力，下一层把窗口错开半格，让信息跨窗口流动，并像 CNN 一样逐层合并图块形成多尺度特征。这让 Transformer 能胜任检测、分割等密集预测任务。","en":"ViT attends globally, so cost explodes at high resolution. Swin attends inside small windows, shifts the windows by half a window in the next layer so information flows across them, and merges patches layer by layer into multi-scale features like a CNN – making Transformers practical for detection and segmentation."},"paper":{"title":"Swin Transformer: Hierarchical Vision Transformer using Shifted Windows","authors":["Ze Liu","Yutong Lin","Yue Cao","et al."],"firstPublished":"2021-03-25","year":2021,"arxiv":"2103.14030","url":"https://arxiv.org/abs/2103.14030","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["vit"],"usedBy":[],"related":["sliding-window","detr"],"usedIn":{"zh":"检测、分割和视频理解的骨干网络","en":"Backbone for detection, segmentation and video understanding"},"resources":[{"title":"microsoft/Swin-Transformer（官方实现）","url":"https://github.com/microsoft/Swin-Transformer","type":"repo","lang":"en"},{"title":"Hugging Face timm 文档（PyTorch 图像模型库）","url":"https://huggingface.co/docs/timm/index","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/swin/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/swin/"}},{"id":"dinov2","name":{"zh":"DINOv2 自监督视觉特征","en":"DINOv2 (Self-Supervised Visual Features)"},"aka":["DINO"],"category":"world-models","tier":"core","year":2023,"difficulty":3,"summary":{"zh":"不用标注，用自蒸馏在海量图片上训练通用视觉特征。","en":"Label-free self-distillation on curated images yields general-purpose visual features."},"explain":{"zh":"学生网络和教师网络看同一张图的不同裁剪，学生要预测教师的输出，教师是学生参数的滑动平均。DINOv2 在精心筛选的 1.42 亿张图片上训练，得到的特征不用微调就能直接做分类、分割、深度估计和检索，是常用的“冻结视觉骨干”。","en":"A student and a teacher network see different crops of the same image; the student predicts the teacher’s output, and the teacher is a moving average of the student. Trained on 142M curated images, DINOv2 features work off the shelf – no fine-tuning – for classification, segmentation, depth and retrieval."},"paper":{"title":"DINOv2: Learning Robust Visual Features without Supervision","authors":["Maxime Oquab","Timothée Darcet","Théo Moutakanni","et al."],"firstPublished":"2023-04-14","year":2023,"arxiv":"2304.07193","url":"https://arxiv.org/abs/2304.07193","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["vit","simclr"],"usedBy":[],"related":["mae","clip","jepa"],"usedIn":{"zh":"Depth Anything 等模型的视觉编码器；图像检索与分割的冻结骨干","en":"Visual encoder in Depth Anything and others; a frozen backbone for retrieval and segmentation"},"resources":[{"title":"facebookresearch/dinov2（官方实现）","url":"https://github.com/facebookresearch/dinov2","type":"repo","lang":"en"},{"title":"Hugging Face 计算机视觉课程","url":"https://huggingface.co/learn/computer-vision-course/unit0/welcome/welcome","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/dinov2/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/dinov2/"}},{"id":"a3c","name":{"zh":"Actor-Critic 与 A3C/A2C","en":"Actor-Critic & A3C/A2C"},"aka":["A2C","Advantage Actor-Critic"],"category":"rl","tier":"core","year":2016,"difficulty":3,"summary":{"zh":"策略网络负责行动、价值网络负责打分，用优势函数降低方差。","en":"An actor picks actions, a critic scores them; the advantage cuts gradient variance."},"explain":{"zh":"纯策略梯度用整局回报当信号，噪声很大。Actor-Critic 加一个价值网络估计“这个状态本来值多少分”，用“实际 − 预期”（优势）来更新策略，学习更稳。A3C 让多个环境副本异步并行采样，后来的同步版本 A2C 更常用，PPO 也沿用了这个框架。","en":"Plain policy gradients use the whole episode’s return, which is noisy. Actor-critic adds a value network that estimates what a state is worth and updates the policy with “actual minus expected” (the advantage), making learning steadier. A3C runs many environment copies asynchronously; its synchronous variant A2C is more common, and PPO keeps the same frame."},"paper":{"title":"Asynchronous Methods for Deep Reinforcement Learning","authors":["Volodymyr Mnih","Adrià Puigdomènech Badia","Mehdi Mirza","et al."],"firstPublished":"2016-02-04","year":2016,"arxiv":"1602.01783","url":"https://arxiv.org/abs/1602.01783","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["policy-gradient","dqn"],"usedBy":[],"related":["ppo","sac"],"usedIn":{"zh":"Stable-Baselines3 等库中的标准算法（A2C）；PPO 的基础框架","en":"A standard algorithm (A2C) in Stable-Baselines3 and other libraries; the frame PPO builds on"},"resources":[{"title":"Lilian Weng：Policy Gradient Algorithms","url":"https://lilianweng.github.io/posts/2018-04-08-policy-gradient/","type":"blog","lang":"en"},{"title":"UC Berkeley CS285 深度强化学习","url":"https://rail.eecs.berkeley.edu/deeprlcourse/","type":"course","lang":"en"},{"title":"Stable-Baselines3 文档","url":"https://stable-baselines3.readthedocs.io/en/master/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/a3c/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/a3c/"}},{"id":"trpo","name":{"zh":"信赖域策略优化（TRPO）","en":"Trust Region Policy Optimization (TRPO)"},"aka":[],"category":"rl","tier":"extended","year":2015,"difficulty":4,"summary":{"zh":"限制每次策略更新的 KL 步长，保证性能单调提升。","en":"Caps each policy update by a KL “trust region” for monotonic improvement."},"explain":{"zh":"策略梯度步子迈大了，策略可能一下子崩掉。TRPO 要求新旧策略的 KL 散度不超过一个阈值，在这个“信赖域”内最大化改进，并用共轭梯度近似求解。它理论漂亮但实现复杂，PPO 用一个裁剪目标近似了同样的思想。","en":"A large policy-gradient step can wreck the policy. TRPO maximizes improvement subject to a cap on KL divergence between old and new policies – a trust region – solved approximately with conjugate gradients. Elegant but complex; PPO approximates the same idea with a clipped objective."},"paper":{"title":"Trust Region Policy Optimization","authors":["John Schulman","Sergey Levine","Philipp Moritz","et al."],"firstPublished":"2015-02-19","year":2015,"arxiv":"1502.05477","url":"https://arxiv.org/abs/1502.05477","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["policy-gradient"],"usedBy":[],"related":["ppo","a3c"],"usedIn":{"zh":"PPO 的直接前身；连续控制基准","en":"The direct predecessor of PPO; continuous-control benchmarks"},"resources":[{"title":"Spinning Up：TRPO","url":"https://spinningup.openai.com/en/latest/algorithms/trpo.html","type":"docs","lang":"en"},{"title":"Lilian Weng：Policy Gradient Algorithms","url":"https://lilianweng.github.io/posts/2018-04-08-policy-gradient/","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/trpo/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/trpo/"}},{"id":"ddpg-td3","name":{"zh":"DDPG / TD3（连续动作控制）","en":"DDPG / TD3 (Continuous Control)"},"aka":["DDPG","TD3","Twin Delayed DDPG"],"category":"rl","tier":"core","year":2015,"difficulty":3,"summary":{"zh":"把 DQN 的思路搬到连续动作空间，用确定性策略加评论家网络。","en":"Brings DQN-style learning to continuous actions with a deterministic actor and a critic."},"explain":{"zh":"机器人关节角这类连续动作没法像 DQN 那样枚举取最大。DDPG 训练一个直接输出动作的策略网络，再用 Q 网络给它打分并回传梯度，配合经验回放和目标网络。TD3 针对 DDPG 高估 Q 值的问题，加入双 Q 取小、延迟更新和目标平滑，更稳定。","en":"You cannot enumerate continuous actions like joint angles the way DQN does. DDPG trains an actor that outputs an action directly and a Q-network that scores it and passes back gradients, with replay and target networks. TD3 fixes DDPG’s Q over-estimation with twin critics (take the min), delayed updates and target smoothing."},"paper":{"title":"Continuous control with deep reinforcement learning","authors":["Timothy P. Lillicrap","Jonathan J. Hunt","Alexander Pritzel","et al."],"firstPublished":"2015-09-09","year":2015,"arxiv":"1509.02971","url":"https://arxiv.org/abs/1509.02971","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["dqn","policy-gradient"],"usedBy":[],"related":["sac"],"usedIn":{"zh":"机器人与 MuJoCo 连续控制基准","en":"Robotics and MuJoCo continuous-control benchmarks"},"resources":[{"title":"Spinning Up：DDPG","url":"https://spinningup.openai.com/en/latest/algorithms/ddpg.html","type":"docs","lang":"en"},{"title":"Spinning Up：TD3","url":"https://spinningup.openai.com/en/latest/algorithms/td3.html","type":"docs","lang":"en"},{"title":"vwxyzjn/cleanrl（单文件 RL 参考实现）","url":"https://github.com/vwxyzjn/cleanrl","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/ddpg-td3/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/ddpg-td3/"}},{"id":"cql","name":{"zh":"保守 Q 学习（CQL，离线强化学习）","en":"Conservative Q-Learning (CQL, Offline RL)"},"aka":["Offline RL","离线强化学习"],"category":"rl","tier":"extended","year":2020,"difficulty":4,"summary":{"zh":"只用固定数据集学策略，压低没见过的动作的 Q 值以免盲目乐观。","en":"Learns from a fixed dataset by pushing down Q-values of unseen actions."},"explain":{"zh":"离线强化学习不能再和环境交互，只能用历史数据。普通 Q 学习会高估数据里没出现过的动作，然后策略就去选这些“想象中很好”的动作。CQL 在损失里额外压低这些动作的 Q 值，得到一个保守的下界，让策略留在数据覆盖的范围内。","en":"Offline RL learns only from logged data, with no new interaction. Ordinary Q-learning over-estimates actions absent from the data, and the policy then chases those imaginary wins. CQL adds a term that pushes their Q-values down, learning a conservative lower bound that keeps the policy within the data’s support."},"paper":{"title":"Conservative Q-Learning for Offline Reinforcement Learning","authors":["Aviral Kumar","Aurick Zhou","George Tucker","et al."],"firstPublished":"2020-06-08","year":2020,"arxiv":"2006.04779","url":"https://arxiv.org/abs/2006.04779","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["q-learning","dqn"],"usedBy":[],"related":["decision-transformer","sac"],"usedIn":{"zh":"推荐、医疗、机器人等只能用历史数据的场景；D4RL 基准","en":"Settings limited to logged data – recommendation, healthcare, robotics; D4RL benchmarks"},"resources":[{"title":"aviralkumar2907/CQL（官方实现）","url":"https://github.com/aviralkumar2907/CQL","type":"repo","lang":"en"},{"title":"UC Berkeley CS285 深度强化学习","url":"https://rail.eecs.berkeley.edu/deeprlcourse/","type":"course","lang":"en"},{"title":"Minari（离线 RL 数据集，D4RL 后继）","url":"https://minari.farama.org/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/cql/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/cql/"}},{"id":"decision-transformer","name":{"zh":"Decision Transformer","en":"Decision Transformer"},"aka":[],"category":"rl","tier":"extended","year":2021,"difficulty":3,"summary":{"zh":"把强化学习当序列建模：给定目标回报，让 Transformer 预测下一个动作。","en":"RL as sequence modelling: condition on a target return and predict the next action."},"explain":{"zh":"把轨迹写成“期望回报、状态、动作”交替的序列，用 GPT 式的因果 Transformer 做监督学习。测试时输入想要的回报，模型就生成相应水平的动作。它不用价值函数和策略梯度，在离线 RL 基准上与当时的专门方法相当。","en":"Write a trajectory as interleaved (return-to-go, state, action) tokens and train a GPT-style causal Transformer with plain supervised learning. At test time, prompt with the return you want and it acts accordingly. No value functions or policy gradients, yet it matched dedicated offline-RL methods of the time."},"paper":{"title":"Decision Transformer: Reinforcement Learning via Sequence Modeling","authors":["Lili Chen","Kevin Lu","Aravind Rajeswaran","et al."],"firstPublished":"2021-06-02","year":2021,"arxiv":"2106.01345","url":"https://arxiv.org/abs/2106.01345","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["transformer","dqn"],"usedBy":[],"related":["cql","world-models"],"usedIn":{"zh":"离线 RL 与“用序列模型做控制”的研究基线","en":"A research baseline for offline RL and sequence-model control"},"resources":[{"title":"kzl/decision-transformer（官方实现）","url":"https://github.com/kzl/decision-transformer","type":"repo","lang":"en"},{"title":"Hugging Face 博客：Decision Transformers","url":"https://huggingface.co/blog/decision-transformers","type":"blog","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/decision-transformer/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/decision-transformer/"}},{"id":"self-refine","name":{"zh":"Self-Refine 自我反馈迭代","en":"Self-Refine"},"aka":[],"category":"reasoning","tier":"core","year":2023,"difficulty":2,"summary":{"zh":"同一个模型先写初稿，再自己点评，再按点评修改，循环几轮。","en":"One model drafts, critiques its own output, then revises – repeated a few times."},"explain":{"zh":"不需要额外训练和外部奖励：让模型生成答案，再用提示让它指出问题，然后带着反馈重写，直到满意或达到轮数上限。在代码优化、写作、数学等 7 类任务上都比单次生成更好，是很多 Agent“反思”步骤的原型。","en":"No extra training or external reward: the model answers, is prompted to critique the answer, then rewrites using the feedback, until it is satisfied or hits a round limit. It beat one-shot generation on seven task types including code, writing and maths, and is the template for many agents’ “reflect” step."},"paper":{"title":"Self-Refine: Iterative Refinement with Self-Feedback","authors":["Aman Madaan","Niket Tandon","Prakhar Gupta","et al."],"firstPublished":"2023-03-30","year":2023,"arxiv":"2303.17651","url":"https://arxiv.org/abs/2303.17651","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["cot"],"usedBy":[],"related":["reflexion","llm-judge"],"usedIn":{"zh":"代码与写作的多轮自我修订；Agent 的反思与自检步骤","en":"Multi-round self-revision of code and writing; agents’ reflection and self-check steps"},"resources":[{"title":"Self-Refine 项目页","url":"https://selfrefine.info/","type":"docs","lang":"en"},{"title":"madaan/self-refine（官方实现）","url":"https://github.com/madaan/self-refine","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/self-refine/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/self-refine/"}},{"id":"normalizing-flows","name":{"zh":"标准化流（RealNVP / Glow）","en":"Normalizing Flows (RealNVP / Glow)"},"aka":["RealNVP","Glow","Normalizing Flow"],"category":"generative","tier":"extended","year":2016,"difficulty":4,"summary":{"zh":"用一串可逆变换把简单分布变成复杂分布，可以精确算似然。","en":"Chains invertible maps from a simple distribution to a complex one, with exact likelihood."},"explain":{"zh":"从标准高斯出发，经过一系列可逆、雅可比行列式好算的变换得到数据分布；反过来也能把数据映射回高斯并精确计算概率。RealNVP 用“仿射耦合层”让这件事又快又可逆。它在图像质量上被扩散模型超越，但精确似然的思路影响了后来的流匹配。","en":"Start from a standard Gaussian and apply invertible transforms whose Jacobian determinant is cheap; run it backwards to map data to the Gaussian and get exact probabilities. RealNVP’s affine coupling layers make this fast and invertible. Diffusion overtook flows on image quality, but the idea fed into flow matching."},"paper":{"title":"Density estimation using Real NVP","authors":["Laurent Dinh","Jascha Sohl-Dickstein","Samy Bengio"],"firstPublished":"2016-05-27","year":2016,"arxiv":"1605.08803","url":"https://arxiv.org/abs/1605.08803","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["vae"],"usedBy":[],"related":["flow-matching","gan"],"usedIn":{"zh":"精确似然的密度估计、物理模拟采样、变分推断","en":"Exact-likelihood density estimation, physics sampling, variational inference"},"resources":[{"title":"Lilian Weng：Flow-based Deep Generative Models","url":"https://lilianweng.github.io/posts/2018-10-13-flow-models/","type":"blog","lang":"en"},{"title":"Papamakarios 等：Normalizing Flows 综述（arXiv:1912.02762）","url":"https://arxiv.org/abs/1912.02762","type":"paper","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/normalizing-flows/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/normalizing-flows/"}},{"id":"ctc","name":{"zh":"CTC 序列对齐损失","en":"Connectionist Temporal Classification (CTC)"},"aka":[],"category":"multimodal","tier":"core","year":2006,"difficulty":3,"summary":{"zh":"不用逐帧标注，就能训练语音、手写等输入输出长度不一的识别模型。","en":"Trains speech and handwriting recognizers without frame-level alignment."},"explain":{"zh":"一段音频有几百帧，对应的文字只有十几个字，事先不知道哪帧对哪个字。CTC 引入“空白”符号，把所有能折叠成正确文字的对齐路径的概率加起来作为目标，用动态规划高效计算。这样只需要“音频 + 文字”成对数据就能端到端训练。","en":"Hundreds of audio frames map to a dozen characters, and nobody knows which frame is which. CTC adds a blank symbol and sums the probability of every alignment that collapses to the correct text, computed efficiently by dynamic programming – so (audio, transcript) pairs alone suffice for end-to-end training."},"paper":{"title":"Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks","authors":["Alex Graves","Santiago Fernández","Faustino Gomez","Jürgen Schmidhuber"],"venue":"ICML 2006","year":2006,"doi":"10.1145/1143844.1143891","url":"https://doi.org/10.1145/1143844.1143891","source":"Crossref API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["lstm"],"usedBy":[],"related":["whisper","wav2vec2"],"usedIn":{"zh":"语音识别（如 DeepSpeech 2、wav2vec 2.0 微调）、手写与场景文字识别","en":"Speech recognition (e.g. DeepSpeech 2, wav2vec 2.0 fine-tuning), handwriting and scene-text OCR"},"resources":[{"title":"Distill：Sequence Modeling with CTC","url":"https://distill.pub/2017/ctc/","type":"blog","lang":"en"},{"title":"PyTorch 文档：nn.CTCLoss","url":"https://docs.pytorch.org/docs/stable/generated/torch.nn.CTCLoss.html","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/ctc/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/ctc/"}},{"id":"wav2vec2","name":{"zh":"wav2vec 2.0 自监督语音表征","en":"wav2vec 2.0"},"aka":[],"category":"multimodal","tier":"core","year":2020,"difficulty":3,"summary":{"zh":"在无标注语音上做对比学习预训练，少量标注就能做好语音识别。","en":"Contrastive self-supervised pre-training on raw audio; strong ASR from little labelled data."},"explain":{"zh":"先用卷积把波形变成帧特征，再遮住一部分，让 Transformer 从候选里认出被遮住位置的正确量化表示（对比学习）。预训练只用无标注音频，之后用很少的转写数据加 CTC 微调就能得到好的识别效果，论文中只用 10 分钟标注数据也能识别。","en":"A conv encoder turns the waveform into frames, some are masked, and a Transformer must pick the true quantized representation for each masked slot from distractors (contrastive learning). Pre-training needs only unlabelled audio; fine-tuning with a little transcribed speech and CTC then gives strong recognition – the paper reports usable results with just 10 minutes of labels."},"paper":{"title":"wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations","authors":["Alexei Baevski","Henry Zhou","Abdelrahman Mohamed","et al."],"firstPublished":"2020-06-20","year":2020,"arxiv":"2006.11477","url":"https://arxiv.org/abs/2006.11477","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["transformer","contrastive-learning"],"usedBy":[],"related":["whisper","ctc"],"usedIn":{"zh":"低资源语言语音识别、语音特征提取（Hugging Face 常用 ASR 模型）","en":"Low-resource speech recognition and speech features (a common ASR model on Hugging Face)"},"resources":[{"title":"Hugging Face Transformers：Wav2Vec2","url":"https://huggingface.co/docs/transformers/model_doc/wav2vec2","type":"docs","lang":"en"},{"title":"fairseq：wav2vec 2.0 官方示例","url":"https://github.com/facebookresearch/fairseq/tree/main/examples/wav2vec","type":"repo","lang":"en"},{"title":"Hugging Face 音频课程","url":"https://huggingface.co/learn/audio-course/chapter0/introduction","type":"course","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/wav2vec2/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/wav2vec2/"}},{"id":"valle","name":{"zh":"VALL-E 神经编解码语言模型 TTS","en":"VALL-E (Neural Codec Language Model TTS)"},"aka":["Neural Codec LM","语音 token"],"category":"multimodal","tier":"extended","year":2023,"difficulty":3,"summary":{"zh":"把语音转成离散编码，用语言模型生成，几秒录音就能模仿音色。","en":"Speech as discrete codec tokens generated by a language model – clones a voice from seconds of audio."},"explain":{"zh":"先用神经音频编解码器把语音变成离散 token，再把 TTS 当成“根据文字和一段提示音频续写 token”的语言建模问题。论文用 3 秒录音作为提示就能合成该说话人的声音，并保留情绪和录音环境。这条“编解码 token + 语言模型”路线成了此后语音生成的主流。","en":"A neural audio codec turns speech into discrete tokens, and TTS becomes language modelling: continue the tokens given text plus a short audio prompt. The paper clones a speaker from a 3-second recording, keeping emotion and acoustic environment. The codec-tokens-plus-LM recipe became the mainstream for speech generation."},"paper":{"title":"Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers","authors":["Chengyi Wang","Sanyuan Chen","Yu Wu","et al."],"firstPublished":"2023-01-05","year":2023,"arxiv":"2301.02111","url":"https://arxiv.org/abs/2301.02111","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["vq-vae","gpt3-icl"],"usedBy":[],"related":["whisper"],"usedIn":{"zh":"零样本语音克隆与 TTS；“编解码 token + LM”式语音生成","en":"Zero-shot voice cloning and TTS; codec-token + LM speech generation"},"resources":[{"title":"Microsoft Research：VALL-E 项目页","url":"https://www.microsoft.com/en-us/research/project/vall-e-x/","type":"docs","lang":"en"},{"title":"facebookresearch/encodec（神经音频编解码器）","url":"https://github.com/facebookresearch/encodec","type":"repo","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/valle/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/valle/"}},{"id":"cross-encoder-rerank","name":{"zh":"交叉编码器重排（BERT Reranker）","en":"Cross-Encoder Reranking"},"aka":["Reranker","monoBERT","重排序"],"category":"retrieval","tier":"core","year":2019,"difficulty":2,"summary":{"zh":"把查询和文档拼在一起送进 BERT 打相关分，给初筛结果重新排序。","en":"Feed query and document together into BERT to score relevance and re-rank candidates."},"explain":{"zh":"向量检索把查询和文档分别编码，快但粗。交叉编码器把两者拼成一个输入，让每一层注意力都能看到双方，所以判断相关性更准，但太慢，只能对前几十到几百条候选重排。“向量/BM25 召回 + 交叉编码器重排”是 RAG 的标准两阶段结构。","en":"Vector search encodes query and document separately – fast but coarse. A cross-encoder concatenates them so every attention layer sees both, judging relevance far more accurately; it is too slow for the whole corpus, so it re-ranks the top tens to hundreds. Retrieve with vectors/BM25, then re-rank: the standard two-stage RAG setup."},"paper":{"title":"Passage Re-ranking with BERT","authors":["Rodrigo Nogueira","Kyunghyun Cho"],"firstPublished":"2019-01-13","year":2019,"arxiv":"1901.04085","url":"https://arxiv.org/abs/1901.04085","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["bert"],"usedBy":[],"related":["colbert","dpr","rag"],"usedIn":{"zh":"RAG 和搜索系统的重排阶段（如 bge-reranker、各家 Rerank API）","en":"The re-ranking stage of RAG and search (e.g. bge-reranker and vendor Rerank APIs)"},"resources":[{"title":"Sentence Transformers：Cross-Encoder 用法","url":"https://sbert.net/docs/cross_encoder/usage/usage.html","type":"docs","lang":"en"},{"title":"Sentence Transformers 文档","url":"https://sbert.net/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/cross-encoder-rerank/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/cross-encoder-rerank/"}},{"id":"autogen","name":{"zh":"多智能体对话框架（AutoGen）","en":"Multi-Agent Conversation (AutoGen)"},"aka":["Multi-Agent","多智能体"],"category":"agents","tier":"core","year":2023,"difficulty":2,"summary":{"zh":"让多个分工不同的 LLM 智能体互相对话、调用工具，协作完成任务。","en":"Multiple specialised LLM agents converse and call tools to solve a task together."},"explain":{"zh":"给每个智能体设定角色（写代码的、审查的、执行的、代表人类的），任务通过它们之间的消息往来推进，需要时调用工具或让人类介入。它把“多智能体协作”变成了可编程的对话模式，是此后许多多智能体框架的参考。","en":"Give each agent a role – coder, reviewer, executor, human proxy – and let the task progress through their messages, calling tools or a human when needed. It turned multi-agent collaboration into programmable conversation patterns and became a reference for later multi-agent frameworks."},"paper":{"title":"AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation","authors":["Qingyun Wu","Gagan Bansal","Jieyu Zhang","et al."],"firstPublished":"2023-08-16","year":2023,"arxiv":"2308.08155","url":"https://arxiv.org/abs/2308.08155","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["react","function-calling"],"usedBy":[],"related":["generative-agents","mcp"],"usedIn":{"zh":"多智能体编程、数据分析与研究助理原型","en":"Multi-agent coding, data analysis and research-assistant prototypes"},"resources":[{"title":"microsoft/autogen","url":"https://github.com/microsoft/autogen","type":"repo","lang":"en"},{"title":"AutoGen 官方文档","url":"https://microsoft.github.io/autogen/stable/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/autogen/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/autogen/"}},{"id":"memgpt","name":{"zh":"MemGPT 分层记忆","en":"MemGPT (Hierarchical Agent Memory)"},"aka":["Letta","Agent Memory","长期记忆"],"category":"agents","tier":"core","year":2023,"difficulty":3,"summary":{"zh":"像操作系统管理内存一样，让 LLM 自己在上下文和外部存储之间搬运记忆。","en":"LLM manages its own memory like an OS – paging between the context window and external storage."},"explain":{"zh":"上下文窗口有限，长对话和大文档放不下。MemGPT 把上下文当“内存”，把数据库当“硬盘”，让模型通过函数调用自己决定什么时候写入、检索、汇总记忆。这样智能体能记住很久以前的对话，也能处理超出窗口的文档。","en":"Context windows are finite, so long chats and large documents do not fit. MemGPT treats the context as RAM and a database as disk, and lets the model decide via function calls when to store, retrieve or summarise memories – so an agent can remember old conversations and work over documents larger than its window."},"paper":{"title":"MemGPT: Towards LLMs as Operating Systems","authors":["Charles Packer","Sarah Wooders","Kevin Lin","et al."],"firstPublished":"2023-10-12","year":2023,"arxiv":"2310.08560","url":"https://arxiv.org/abs/2310.08560","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["rag","function-calling"],"usedBy":[],"related":["generative-agents","autogen"],"usedIn":{"zh":"有长期记忆的对话 Agent（Letta 框架的前身）","en":"Chat agents with long-term memory (the precursor of the Letta framework)"},"resources":[{"title":"letta-ai/letta（原 MemGPT）","url":"https://github.com/letta-ai/letta","type":"repo","lang":"en"},{"title":"MemGPT 研究项目页","url":"https://research.memgpt.ai/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/memgpt/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/memgpt/"}},{"id":"membership-inference","name":{"zh":"成员推断攻击","en":"Membership Inference Attacks"},"aka":[],"category":"safety","tier":"core","year":2016,"difficulty":3,"summary":{"zh":"判断某条数据是否被用来训练过模型，用来衡量隐私泄露。","en":"Tells whether a given record was in a model’s training set – a measure of privacy leakage."},"explain":{"zh":"模型往往对训练过的样本更“自信”。攻击者训练若干模仿目标模型的“影子模型”，学会根据输出置信度区分“见过”和“没见过”的样本。它是评估模型隐私风险的标准方法，也用来检测大模型是否背过版权文本或测试集。差分隐私训练（DP-SGD）是主要防御。","en":"Models tend to be more confident on examples they were trained on. The attacker trains shadow models that mimic the target and learns to tell “seen” from “unseen” records by the output confidence. It is the standard way to audit privacy risk, also used to probe whether LLMs memorised copyrighted text or test sets; DP-SGD is the main defence."},"paper":{"title":"Membership Inference Attacks against Machine Learning Models","authors":["Reza Shokri","Marco Stronati","Congzheng Song","et al."],"firstPublished":"2016-10-18","year":2016,"arxiv":"1610.05820","url":"https://arxiv.org/abs/1610.05820","source":"arXiv API","check":"verified"},"verify":{"status":"verified","checkedOn":"2026-10-22"},"prereqs":["logistic-regression"],"usedBy":[],"related":["dp-sgd","watermarking"],"usedIn":{"zh":"模型隐私审计、训练数据泄露与测试集污染检测","en":"Privacy audits, training-data leakage and test-set contamination checks"},"resources":[{"title":"ML Privacy Meter（成员推断审计工具）","url":"https://github.com/privacytrustlab/ml_privacy_meter","type":"repo","lang":"en"},{"title":"Opacus（PyTorch 差分隐私库）","url":"https://opacus.ai/","type":"docs","lang":"en"}],"pages":{"en":"https://aimap.page/en/learn/algorithms/membership-inference/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/membership-inference/"}}],"paths":[{"id":"p-ml-to-dl","title":{"zh":"从零到深度学习","en":"Zero to Deep Learning"},"level":1,"steps":[["linear-regression","logistic-regression","sgd"],["decision-tree","random-forest","gbdt","xgboost","svm","knn","kmeans","pca"],["perceptron","backprop","adam","adamw","dropout","batchnorm","layernorm"],["cnn","resnet","lstm","seq2seq","word2vec","bahdanau-attention"]],"units":["u20-ml-workflow","u21-trees","u30-nn-backprop","u31-training-tricks","u32-cnn-rnn"],"pages":{"en":"https://aimap.page/en/learn/algorithms/paths/p-ml-to-dl/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/paths/p-ml-to-dl/"}},{"id":"p-build-llm","title":{"zh":"从零理解一个大模型","en":"Understanding an LLM End to End"},"level":2,"steps":[["bpe","word2vec"],["transformer","gpt3-icl","bert","rope","rmsnorm","swiglu"],["gqa","mla","flashattention","kv-cache","sampling"],["scaling-laws","chinchilla","sparse-moe","deepseek-moe","mixed-precision","zero"]],"units":["u40-tokens","u41-transformer","u42-modern-arch","u44-pretraining","u60-decoding"],"pages":{"en":"https://aimap.page/en/learn/algorithms/paths/p-build-llm/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/paths/p-build-llm/"}},{"id":"p-post-training","title":{"zh":"后训练与推理模型","en":"Post-training & Reasoning Models"},"level":3,"steps":[["instruction-tuning","self-instruct","lora","qlora"],["policy-gradient","ppo","reward-model"],["rlhf","dpo","simpo","constitutional-ai"],["cot","self-consistency","prm","grpo","rlvr","deepseek-r1","test-time-scaling","dapo"]],"units":["u50-sft-peft","u33-rl-basics","u51-alignment","u52-reasoning"],"pages":{"en":"https://aimap.page/en/learn/algorithms/paths/p-post-training/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/paths/p-post-training/"}},{"id":"p-ai-engineer","title":{"zh":"AI 应用工程师","en":"AI Application Engineer"},"level":2,"steps":[["bm25","sentence-bert","dpr","hnsw","product-quantization"],["rag","hyde","colbert","graphrag"],["react","function-calling","mcp","reflexion","swe-agent"],["paged-attention","continuous-batching","speculative-decoding","awq","llm-judge","prompt-injection"]],"units":["u80-retrieval","u81-rag","u82-agents","u62-serving","u83-ai-engineering","u92-security"],"pages":{"en":"https://aimap.page/en/learn/algorithms/paths/p-ai-engineer/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/paths/p-ai-engineer/"}},{"id":"p-frontier","title":{"zh":"生成式与前沿研究","en":"Generative Models & Frontier Research"},"level":4,"steps":[["vae","ddpm","latent-diffusion","cfg","dit","flow-matching","rectified-flow"],["vit","clip","llava","whisper","sam"],["mamba","linear-attention","gated-deltanet","jamba","diffusion-lm","nsa"],["world-models","dreamer","jepa","genie"],["superposition","interp-sae","circuit-tracing","activation-steering"]],"units":["u72-diffusion","u71-vlm","u43-moe-newarch","u73-speech-video-world","u91-interpretability"],"pages":{"en":"https://aimap.page/en/learn/algorithms/paths/p-frontier/","zh-hans":"https://aimap.page/zh-hans/learn/algorithms/paths/p-frontier/"}}],"curriculum":{"stages":[{"id":"s0-literacy","order":0,"title":{"zh":"预备：AI 通识与上手","en":"Prep: AI Literacy"},"units":["u00-what-is-ai","u01-prompting","u02-hallucination"],"hours":8,"pages":{"en":"https://aimap.page/en/learn/stage/s0-literacy/","zh-hans":"https://aimap.page/zh-hans/learn/stage/s0-literacy/"}},{"id":"s1-math-code","order":1,"title":{"zh":"数学与编程基础","en":"Math & Programming Foundations"},"units":["u10-python","u11-linalg","u12-calculus","u13-probability","u14-pytorch"],"hours":61,"pages":{"en":"https://aimap.page/en/learn/stage/s1-math-code/","zh-hans":"https://aimap.page/zh-hans/learn/stage/s1-math-code/"}},{"id":"s2-classic-ml","order":2,"title":{"zh":"经典机器学习","en":"Classical Machine Learning"},"units":["u20-ml-workflow","u21-trees","u22-svm-knn","u23-unsupervised"],"hours":32,"pages":{"en":"https://aimap.page/en/learn/stage/s2-classic-ml/","zh-hans":"https://aimap.page/zh-hans/learn/stage/s2-classic-ml/"}},{"id":"s3-deep-learning","order":3,"title":{"zh":"深度学习","en":"Deep Learning"},"units":["u30-nn-backprop","u31-training-tricks","u32-cnn-rnn","u33-rl-basics"],"hours":44,"pages":{"en":"https://aimap.page/en/learn/stage/s3-deep-learning/","zh-hans":"https://aimap.page/zh-hans/learn/stage/s3-deep-learning/"}},{"id":"s4-transformer-llm","order":4,"title":{"zh":"Transformer 与大模型","en":"Transformers & LLMs"},"units":["u40-tokens","u41-transformer","u42-modern-arch","u43-moe-newarch","u44-pretraining"],"hours":54,"pages":{"en":"https://aimap.page/en/learn/stage/s4-transformer-llm/","zh-hans":"https://aimap.page/zh-hans/learn/stage/s4-transformer-llm/"}},{"id":"s5-training-alignment","order":5,"title":{"zh":"训练、对齐与推理模型","en":"Post-training, Alignment & Reasoning Models"},"units":["u50-sft-peft","u51-alignment","u52-reasoning"],"hours":32,"pages":{"en":"https://aimap.page/en/learn/stage/s5-training-alignment/","zh-hans":"https://aimap.page/zh-hans/learn/stage/s5-training-alignment/"}},{"id":"s6-inference-efficiency","order":6,"title":{"zh":"推理部署与效率","en":"Inference, Deployment & Efficiency"},"units":["u60-decoding","u61-compression","u62-serving","u63-local","u64-systems"],"hours":36,"pages":{"en":"https://aimap.page/en/learn/stage/s6-inference-efficiency/","zh-hans":"https://aimap.page/zh-hans/learn/stage/s6-inference-efficiency/"}},{"id":"s7-multimodal-gen","order":7,"title":{"zh":"多模态与生成","en":"Multimodal & Generative AI"},"units":["u70-vision","u71-vlm","u72-diffusion","u73-speech-video-world"],"hours":38,"pages":{"en":"https://aimap.page/en/learn/stage/s7-multimodal-gen/","zh-hans":"https://aimap.page/zh-hans/learn/stage/s7-multimodal-gen/"}},{"id":"s8-rag-agents","order":8,"title":{"zh":"RAG、Agent 与应用工程","en":"RAG, Agents & AI Engineering"},"units":["u80-retrieval","u81-rag","u82-agents","u83-ai-engineering"],"hours":36,"pages":{"en":"https://aimap.page/en/learn/stage/s8-rag-agents/","zh-hans":"https://aimap.page/zh-hans/learn/stage/s8-rag-agents/"}},{"id":"s9-safety-frontier","order":9,"title":{"zh":"安全、评测与前沿","en":"Safety, Evaluation & Frontier"},"units":["u90-evaluation","u91-interpretability","u92-security","u93-alignment-governance","u94-ai4science","u95-frontier"],"hours":44,"pages":{"en":"https://aimap.page/en/learn/stage/s9-safety-frontier/","zh-hans":"https://aimap.page/zh-hans/learn/stage/s9-safety-frontier/"}}],"units":[{"id":"u00-what-is-ai","stage":"s0-literacy","title":{"zh":"AI、机器学习、深度学习与大模型是什么","en":"What AI, ML, Deep Learning and LLMs Are"},"hours":3,"algorithms":["gpt3-icl","transformer"],"prereqUnits":[],"papers":[{"ref":"arxiv:2005.14165","title":"Language Models are Few-Shot Learners","year":2020,"url":"https://arxiv.org/abs/2005.14165","classicsId":"arxiv-2005.14165","check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u00-what-is-ai/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u00-what-is-ai/"}},{"id":"u01-prompting","stage":"s0-literacy","title":{"zh":"提示词与高效使用 AI","en":"Prompting & Using AI Well"},"hours":3,"algorithms":["cot","gpt3-icl","self-consistency"],"prereqUnits":["u00-what-is-ai"],"papers":[{"ref":"arxiv:2201.11903","title":"Chain-of-Thought Prompting Elicits Reasoning in Large Language Models","year":2022,"url":"https://arxiv.org/abs/2201.11903","classicsId":"arxiv-2201.11903","check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u01-prompting/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u01-prompting/"}},{"id":"u02-hallucination","stage":"s0-literacy","title":{"zh":"幻觉、可靠性与使用边界","en":"Hallucination, Reliability & Limits"},"hours":2,"algorithms":["rag","sampling"],"prereqUnits":["u00-what-is-ai"],"papers":[{"ref":"arxiv:2005.11401","title":"Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks","year":2020,"url":"https://arxiv.org/abs/2005.11401","classicsId":"arxiv-2005.11401","check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u02-hallucination/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u02-hallucination/"}},{"id":"u10-python","stage":"s1-math-code","title":{"zh":"Python、NumPy 与开发环境","en":"Python, NumPy & Tooling"},"hours":12,"algorithms":[],"prereqUnits":[],"papers":[],"pages":{"en":"https://aimap.page/en/learn/unit/u10-python/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u10-python/"}},{"id":"u11-linalg","stage":"s1-math-code","title":{"zh":"线性代数：向量、矩阵与分解","en":"Linear Algebra"},"hours":15,"algorithms":["pca","lora"],"prereqUnits":["u10-python"],"papers":[],"pages":{"en":"https://aimap.page/en/learn/unit/u11-linalg/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u11-linalg/"}},{"id":"u12-calculus","stage":"s1-math-code","title":{"zh":"微积分与优化","en":"Calculus & Optimization"},"hours":12,"algorithms":["sgd","backprop"],"prereqUnits":["u10-python"],"papers":[{"ref":"doi:10.1214/aoms/1177729586","title":"A Stochastic Approximation Method","year":1951,"url":"https://doi.org/10.1214/aoms/1177729586","classicsId":null,"check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u12-calculus/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u12-calculus/"}},{"id":"u13-probability","stage":"s1-math-code","title":{"zh":"概率、统计与信息论","en":"Probability, Statistics & Information Theory"},"hours":12,"algorithms":["logistic-regression","em","vae"],"prereqUnits":["u10-python"],"papers":[],"pages":{"en":"https://aimap.page/en/learn/unit/u13-probability/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u13-probability/"}},{"id":"u14-pytorch","stage":"s1-math-code","title":{"zh":"PyTorch 入门与自动求导","en":"PyTorch & Autograd"},"hours":10,"algorithms":["backprop","adam"],"prereqUnits":["u10-python","u12-calculus"],"papers":[],"pages":{"en":"https://aimap.page/en/learn/unit/u14-pytorch/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u14-pytorch/"}},{"id":"u20-ml-workflow","stage":"s2-classic-ml","title":{"zh":"机器学习流程：数据、损失与泛化","en":"The ML Workflow: Data, Loss & Generalization"},"hours":10,"algorithms":["linear-regression","logistic-regression","sgd"],"prereqUnits":["u12-calculus","u13-probability"],"papers":[{"ref":"arxiv:1811.12808","title":"Model Evaluation, Model Selection, and Algorithm Selection in Machine Learning","year":2018,"url":"https://arxiv.org/abs/1811.12808","classicsId":null,"check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u20-ml-workflow/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u20-ml-workflow/"}},{"id":"u21-trees","stage":"s2-classic-ml","title":{"zh":"树模型与集成学习","en":"Trees & Ensembles"},"hours":8,"algorithms":["decision-tree","random-forest","gbdt","xgboost","lightgbm","adaboost"],"prereqUnits":["u20-ml-workflow"],"papers":[{"ref":"doi:10.1023/A:1010933404324","title":"Random Forests","year":2001,"url":"https://doi.org/10.1023/A:1010933404324","classicsId":null,"check":"verified"},{"ref":"doi:10.1214/aos/1013203451","title":"Greedy function approximation: A gradient boosting machine.","year":2001,"url":"https://doi.org/10.1214/aos/1013203451","classicsId":null,"check":"verified"},{"ref":"arxiv:1603.02754","title":"XGBoost: A Scalable Tree Boosting System","year":2016,"url":"https://arxiv.org/abs/1603.02754","classicsId":null,"check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u21-trees/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u21-trees/"}},{"id":"u22-svm-knn","stage":"s2-classic-ml","title":{"zh":"线性模型、SVM 与近邻","en":"Linear Models, SVM & Nearest Neighbors"},"hours":6,"algorithms":["perceptron","svm","knn","lasso","naive-bayes"],"prereqUnits":["u20-ml-workflow","u11-linalg"],"papers":[{"ref":"doi:10.1007/BF00994018","title":"Support-vector networks","year":1995,"url":"https://doi.org/10.1007/BF00994018","classicsId":null,"check":"verified"},{"ref":"doi:10.1109/TIT.1967.1053964","title":"Nearest neighbor pattern classification","year":1967,"url":"https://doi.org/10.1109/TIT.1967.1053964","classicsId":null,"check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u22-svm-knn/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u22-svm-knn/"}},{"id":"u23-unsupervised","stage":"s2-classic-ml","title":{"zh":"无监督学习：聚类、降维与 EM","en":"Unsupervised Learning"},"hours":8,"algorithms":["kmeans","em","pca","tsne","umap","dbscan"],"prereqUnits":["u20-ml-workflow","u11-linalg"],"papers":[{"ref":"doi:10.1111/j.2517-6161.1977.tb01600.x","title":"Maximum Likelihood from Incomplete Data Via the EM Algorithm","year":1977,"url":"https://doi.org/10.1111/j.2517-6161.1977.tb01600.x","classicsId":null,"check":"verified"},{"ref":"arxiv:1802.03426","title":"UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction","year":2018,"url":"https://arxiv.org/abs/1802.03426","classicsId":null,"check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u23-unsupervised/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u23-unsupervised/"}},{"id":"u30-nn-backprop","stage":"s3-deep-learning","title":{"zh":"神经网络与反向传播","en":"Neural Networks & Backpropagation"},"hours":10,"algorithms":["perceptron","backprop","gelu","autoencoder"],"prereqUnits":["u14-pytorch","u20-ml-workflow"],"papers":[{"ref":"doi:10.1037/h0042519","title":"The perceptron: A probabilistic model for information storage and organization in the brain.","year":1958,"url":"https://doi.org/10.1037/h0042519","classicsId":null,"check":"verified"},{"ref":"doi:10.1038/323533a0","title":"Learning representations by back-propagating errors","year":1986,"url":"https://doi.org/10.1038/323533a0","classicsId":"doi-10-1038-323533a0","check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u30-nn-backprop/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u30-nn-backprop/"}},{"id":"u31-training-tricks","stage":"s3-deep-learning","title":{"zh":"训练深层网络：优化器、归一化与正则","en":"Training Deep Nets"},"hours":10,"algorithms":["sgd","adam","adamw","muon","batchnorm","layernorm","rmsnorm","dropout","resnet","weight-init"],"prereqUnits":["u30-nn-backprop"],"papers":[{"ref":"arxiv:1412.6980","title":"Adam: A Method for Stochastic Optimization","year":2014,"url":"https://arxiv.org/abs/1412.6980","classicsId":"arxiv-1412.6980","check":"verified"},{"ref":"arxiv:1711.05101","title":"Decoupled Weight Decay Regularization","year":2017,"url":"https://arxiv.org/abs/1711.05101","classicsId":null,"check":"verified"},{"ref":"arxiv:1502.03167","title":"Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift","year":2015,"url":"https://arxiv.org/abs/1502.03167","classicsId":"arxiv-1502.03167","check":"verified"},{"ref":"arxiv:1512.03385","title":"Deep Residual Learning for Image Recognition","year":2015,"url":"https://arxiv.org/abs/1512.03385","classicsId":"arxiv-1512.03385","check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u31-training-tricks/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u31-training-tricks/"}},{"id":"u32-cnn-rnn","stage":"s3-deep-learning","title":{"zh":"CNN、RNN 与注意力的诞生","en":"CNNs, RNNs and the Birth of Attention"},"hours":10,"algorithms":["cnn","resnet","lstm","seq2seq","word2vec","bahdanau-attention","gru"],"prereqUnits":["u31-training-tricks"],"papers":[{"ref":"doi:10.1109/5.726791","title":"Gradient-based learning applied to document recognition","year":1998,"url":"https://doi.org/10.1109/5.726791","classicsId":"doi-10-1109-5-726791","check":"verified"},{"ref":"doi:10.1162/neco.1997.9.8.1735","title":"Long Short-Term Memory","year":1997,"url":"https://doi.org/10.1162/neco.1997.9.8.1735","classicsId":"doi-10-1162-neco-1997-9-8-1735","check":"verified"},{"ref":"arxiv:1409.3215","title":"Sequence to Sequence Learning with Neural Networks","year":2014,"url":"https://arxiv.org/abs/1409.3215","classicsId":"arxiv-1409.3215","check":"verified"},{"ref":"arxiv:1409.0473","title":"Neural Machine Translation by Jointly Learning to Align and Translate","year":2014,"url":"https://arxiv.org/abs/1409.0473","classicsId":"arxiv-1409.0473","check":"verified"},{"ref":"arxiv:1301.3781","title":"Efficient Estimation of Word Representations in Vector Space","year":2013,"url":"https://arxiv.org/abs/1301.3781","classicsId":"arxiv-1301.3781","check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u32-cnn-rnn/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u32-cnn-rnn/"}},{"id":"u33-rl-basics","stage":"s3-deep-learning","title":{"zh":"强化学习基础","en":"Reinforcement Learning Basics"},"hours":14,"algorithms":["q-learning","dqn","policy-gradient","ppo","sac","mcts-alphazero","muzero","a3c","trpo","ddpg-td3","cql","decision-transformer"],"prereqUnits":["u31-training-tricks","u13-probability"],"papers":[{"ref":"doi:10.1007/BF00992698","title":"Q-learning","year":1992,"url":"https://doi.org/10.1007/BF00992698","classicsId":null,"check":"verified"},{"ref":"arxiv:1312.5602","title":"Playing Atari with Deep Reinforcement Learning","year":2013,"url":"https://arxiv.org/abs/1312.5602","classicsId":"arxiv-1312.5602","check":"verified"},{"ref":"arxiv:1707.06347","title":"Proximal Policy Optimization Algorithms","year":2017,"url":"https://arxiv.org/abs/1707.06347","classicsId":"arxiv-1707.06347","check":"verified"},{"ref":"doi:10.1038/nature16961","title":"Mastering the game of Go with deep neural networks and tree search","year":2016,"url":"https://doi.org/10.1038/nature16961","classicsId":"doi-10-1038-nature16961","check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u33-rl-basics/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u33-rl-basics/"}},{"id":"u40-tokens","stage":"s4-transformer-llm","title":{"zh":"分词与嵌入","en":"Tokenization & Embeddings"},"hours":6,"algorithms":["bpe","word2vec","sentence-bert","sentencepiece","glove","elmo"],"prereqUnits":["u32-cnn-rnn"],"papers":[{"ref":"arxiv:1508.07909","title":"Neural Machine Translation of Rare Words with Subword Units","year":2015,"url":"https://arxiv.org/abs/1508.07909","classicsId":null,"check":"verified"},{"ref":"arxiv:1908.10084","title":"Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks","year":2019,"url":"https://arxiv.org/abs/1908.10084","classicsId":null,"check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u40-tokens/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u40-tokens/"}},{"id":"u41-transformer","stage":"s4-transformer-llm","title":{"zh":"Transformer 与注意力","en":"Transformers & Attention"},"hours":16,"algorithms":["transformer","bert","t5","gpt3-icl","rope","layernorm"],"prereqUnits":["u40-tokens","u31-training-tricks"],"papers":[{"ref":"arxiv:1706.03762","title":"Attention Is All You Need","year":2017,"url":"https://arxiv.org/abs/1706.03762","classicsId":"arxiv-1706.03762","check":"verified"},{"ref":"arxiv:1810.04805","title":"BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding","year":2018,"url":"https://arxiv.org/abs/1810.04805","classicsId":"arxiv-1810.04805","check":"verified"},{"ref":"arxiv:2005.14165","title":"Language Models are Few-Shot Learners","year":2020,"url":"https://arxiv.org/abs/2005.14165","classicsId":"arxiv-2005.14165","check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u41-transformer/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u41-transformer/"}},{"id":"u42-modern-arch","stage":"s4-transformer-llm","title":{"zh":"现代 LLM 架构：位置编码、GQA/MLA 与长上下文","en":"Modern LLM Architecture"},"hours":12,"algorithms":["rope","alibi","swiglu","rmsnorm","mqa","gqa","mla","flashattention","sliding-window","nsa","yarn","ring-attention"],"prereqUnits":["u41-transformer"],"papers":[{"ref":"arxiv:2104.09864","title":"RoFormer: Enhanced Transformer with Rotary Position Embedding","year":2021,"url":"https://arxiv.org/abs/2104.09864","classicsId":"arxiv-2104.09864","check":"verified"},{"ref":"arxiv:2305.13245","title":"GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints","year":2023,"url":"https://arxiv.org/abs/2305.13245","classicsId":"arxiv-2305.13245","check":"verified"},{"ref":"arxiv:2405.04434","title":"DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model","year":2024,"url":"https://arxiv.org/abs/2405.04434","classicsId":"arxiv-2405.04434","check":"verified"},{"ref":"arxiv:2205.14135","title":"FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness","year":2022,"url":"https://arxiv.org/abs/2205.14135","classicsId":"arxiv-2205.14135","check":"verified"},{"ref":"arxiv:2309.00071","title":"YaRN: Efficient Context Window Extension of Large Language Models","year":2023,"url":"https://arxiv.org/abs/2309.00071","classicsId":null,"check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u42-modern-arch/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u42-modern-arch/"}},{"id":"u43-moe-newarch","stage":"s4-transformer-llm","title":{"zh":"MoE 与新架构","en":"MoE & Beyond-Transformer Architectures"},"hours":10,"algorithms":["sparse-moe","gshard","switch-transformer","mixtral","deepseek-moe","aux-free-balancing","s4","mamba","mamba2","linear-attention","rwkv","retnet","hyena","xlstm","gated-deltanet","jamba","diffusion-lm","looped-lm"],"prereqUnits":["u42-modern-arch"],"papers":[{"ref":"arxiv:1701.06538","title":"Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer","year":2017,"url":"https://arxiv.org/abs/1701.06538","classicsId":"arxiv-1701.06538","check":"verified"},{"ref":"arxiv:2101.03961","title":"Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity","year":2021,"url":"https://arxiv.org/abs/2101.03961","classicsId":null,"check":"verified"},{"ref":"arxiv:2401.06066","title":"DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models","year":2024,"url":"https://arxiv.org/abs/2401.06066","classicsId":null,"check":"verified"},{"ref":"arxiv:2312.00752","title":"Mamba: Linear-Time Sequence Modeling with Selective State Spaces","year":2023,"url":"https://arxiv.org/abs/2312.00752","classicsId":"arxiv-2312.00752","check":"verified"},{"ref":"arxiv:2412.06464","title":"Gated Delta Networks: Improving Mamba2 with Delta Rule","year":2024,"url":"https://arxiv.org/abs/2412.06464","classicsId":null,"check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u43-moe-newarch/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u43-moe-newarch/"}},{"id":"u44-pretraining","stage":"s4-transformer-llm","title":{"zh":"预训练、数据与缩放定律","en":"Pretraining, Data & Scaling Laws"},"hours":10,"algorithms":["scaling-laws","chinchilla","emergence","mup","mtp","mixed-precision"],"prereqUnits":["u41-transformer"],"papers":[{"ref":"arxiv:2001.08361","title":"Scaling Laws for Neural Language Models","year":2020,"url":"https://arxiv.org/abs/2001.08361","classicsId":"arxiv-2001.08361","check":"verified"},{"ref":"arxiv:2203.15556","title":"Training Compute-Optimal Large Language Models","year":2022,"url":"https://arxiv.org/abs/2203.15556","classicsId":"arxiv-2203.15556","check":"verified"},{"ref":"arxiv:2206.07682","title":"Emergent Abilities of Large Language Models","year":2022,"url":"https://arxiv.org/abs/2206.07682","classicsId":null,"check":"verified"},{"ref":"arxiv:2404.19737","title":"Better & Faster Large Language Models via Multi-token Prediction","year":2024,"url":"https://arxiv.org/abs/2404.19737","classicsId":null,"check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u44-pretraining/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u44-pretraining/"}},{"id":"u50-sft-peft","stage":"s5-training-alignment","title":{"zh":"指令微调与参数高效微调","en":"Instruction Tuning & PEFT"},"hours":10,"algorithms":["instruction-tuning","self-instruct","adapters","prompt-tuning","lora","qlora","dora","model-merging"],"prereqUnits":["u44-pretraining"],"papers":[{"ref":"arxiv:2109.01652","title":"Finetuned Language Models Are Zero-Shot Learners","year":2021,"url":"https://arxiv.org/abs/2109.01652","classicsId":"arxiv-2109.01652","check":"verified"},{"ref":"arxiv:2106.09685","title":"LoRA: Low-Rank Adaptation of Large Language Models","year":2021,"url":"https://arxiv.org/abs/2106.09685","classicsId":"arxiv-2106.09685","check":"verified"},{"ref":"arxiv:2305.14314","title":"QLoRA: Efficient Finetuning of Quantized LLMs","year":2023,"url":"https://arxiv.org/abs/2305.14314","classicsId":"arxiv-2305.14314","check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u50-sft-peft/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u50-sft-peft/"}},{"id":"u51-alignment","stage":"s5-training-alignment","title":{"zh":"对齐：RLHF、DPO 与宪法式 AI","en":"Alignment: RLHF, DPO & Constitutional AI"},"hours":10,"algorithms":["reward-model","rlhf","dpo","kto","orpo","simpo","constitutional-ai"],"prereqUnits":["u50-sft-peft","u33-rl-basics"],"papers":[{"ref":"arxiv:1706.03741","title":"Deep reinforcement learning from human preferences","year":2017,"url":"https://arxiv.org/abs/1706.03741","classicsId":"arxiv-1706.03741","check":"verified"},{"ref":"arxiv:2203.02155","title":"Training language models to follow instructions with human feedback","year":2022,"url":"https://arxiv.org/abs/2203.02155","classicsId":"arxiv-2203.02155","check":"verified"},{"ref":"arxiv:2305.18290","title":"Direct Preference Optimization: Your Language Model is Secretly a Reward Model","year":2023,"url":"https://arxiv.org/abs/2305.18290","classicsId":"arxiv-2305.18290","check":"verified"},{"ref":"arxiv:2212.08073","title":"Constitutional AI: Harmlessness from AI Feedback","year":2022,"url":"https://arxiv.org/abs/2212.08073","classicsId":"arxiv-2212.08073","check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u51-alignment/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u51-alignment/"}},{"id":"u52-reasoning","stage":"s5-training-alignment","title":{"zh":"推理模型：RLVR、GRPO 与测试时计算","en":"Reasoning Models: RLVR, GRPO & Test-Time Compute"},"hours":12,"algorithms":["cot","self-consistency","tree-of-thoughts","prm","star","grpo","rlvr","dapo","deepseek-r1","test-time-scaling","budget-forcing","self-refine"],"prereqUnits":["u51-alignment"],"papers":[{"ref":"arxiv:2201.11903","title":"Chain-of-Thought Prompting Elicits Reasoning in Large Language Models","year":2022,"url":"https://arxiv.org/abs/2201.11903","classicsId":"arxiv-2201.11903","check":"verified"},{"ref":"arxiv:2305.20050","title":"Let's Verify Step by Step","year":2023,"url":"https://arxiv.org/abs/2305.20050","classicsId":"arxiv-2305.20050","check":"verified"},{"ref":"arxiv:2402.03300","title":"DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models","year":2024,"url":"https://arxiv.org/abs/2402.03300","classicsId":"arxiv-2402.03300","check":"verified"},{"ref":"arxiv:2501.12948","title":"DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning","year":2025,"url":"https://arxiv.org/abs/2501.12948","classicsId":"arxiv-2501.12948","check":"verified"},{"ref":"arxiv:2408.03314","title":"Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters","year":2024,"url":"https://arxiv.org/abs/2408.03314","classicsId":"arxiv-2408.03314","check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u52-reasoning/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u52-reasoning/"}},{"id":"u60-decoding","stage":"s6-inference-efficiency","title":{"zh":"解码、采样与 KV Cache","en":"Decoding, Sampling & KV Cache"},"hours":6,"algorithms":["sampling","kv-cache","streaming-llm","gqa"],"prereqUnits":["u42-modern-arch"],"papers":[{"ref":"arxiv:1904.09751","title":"The Curious Case of Neural Text Degeneration","year":2019,"url":"https://arxiv.org/abs/1904.09751","classicsId":null,"check":"verified"},{"ref":"arxiv:2309.17453","title":"Efficient Streaming Language Models with Attention Sinks","year":2023,"url":"https://arxiv.org/abs/2309.17453","classicsId":null,"check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u60-decoding/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u60-decoding/"}},{"id":"u61-compression","stage":"s6-inference-efficiency","title":{"zh":"量化、剪枝与蒸馏","en":"Quantization, Pruning & Distillation"},"hours":8,"algorithms":["quantization-int8","gptq","awq","smoothquant","sparsegpt","bitnet","knowledge-distillation"],"prereqUnits":["u60-decoding"],"papers":[{"ref":"arxiv:2208.07339","title":"LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale","year":2022,"url":"https://arxiv.org/abs/2208.07339","classicsId":null,"check":"verified"},{"ref":"arxiv:2210.17323","title":"GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers","year":2022,"url":"https://arxiv.org/abs/2210.17323","classicsId":"arxiv-2210.17323","check":"verified"},{"ref":"arxiv:2306.00978","title":"AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration","year":2023,"url":"https://arxiv.org/abs/2306.00978","classicsId":null,"check":"verified"},{"ref":"arxiv:1503.02531","title":"Distilling the Knowledge in a Neural Network","year":2015,"url":"https://arxiv.org/abs/1503.02531","classicsId":"arxiv-1503.02531","check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u61-compression/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u61-compression/"}},{"id":"u62-serving","stage":"s6-inference-efficiency","title":{"zh":"推理引擎与服务","en":"Serving Engines"},"hours":8,"algorithms":["paged-attention","continuous-batching","radix-attention","speculative-decoding","medusa","eagle","pd-disaggregation"],"prereqUnits":["u60-decoding"],"papers":[{"ref":"arxiv:2309.06180","title":"Efficient Memory Management for Large Language Model Serving with PagedAttention","year":2023,"url":"https://arxiv.org/abs/2309.06180","classicsId":"arxiv-2309.06180","check":"verified"},{"ref":"arxiv:2211.17192","title":"Fast Inference from Transformers via Speculative Decoding","year":2022,"url":"https://arxiv.org/abs/2211.17192","classicsId":"arxiv-2211.17192","check":"verified"},{"ref":"arxiv:2312.07104","title":"SGLang: Efficient Execution of Structured Language Model Programs","year":2023,"url":"https://arxiv.org/abs/2312.07104","classicsId":null,"check":"verified"},{"ref":"arxiv:2401.09670","title":"DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving","year":2024,"url":"https://arxiv.org/abs/2401.09670","classicsId":null,"check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u62-serving/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u62-serving/"}},{"id":"u63-local","stage":"s6-inference-efficiency","title":{"zh":"在本地运行大模型","en":"Running LLMs Locally"},"hours":4,"algorithms":["quantization-int8","gptq","kv-cache"],"prereqUnits":["u61-compression"],"papers":[],"pages":{"en":"https://aimap.page/en/learn/unit/u63-local/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u63-local/"}},{"id":"u64-systems","stage":"s6-inference-efficiency","title":{"zh":"GPU、分布式训练与 AI 芯片","en":"GPUs, Distributed Training & AI Chips"},"hours":10,"algorithms":["mixed-precision","zero","tensor-parallel","pipeline-parallel","ring-attention","gshard"],"prereqUnits":["u44-pretraining"],"papers":[{"ref":"arxiv:1710.03740","title":"Mixed Precision Training","year":2017,"url":"https://arxiv.org/abs/1710.03740","classicsId":null,"check":"verified"},{"ref":"arxiv:1910.02054","title":"ZeRO: Memory Optimizations Toward Training Trillion Parameter Models","year":2019,"url":"https://arxiv.org/abs/1910.02054","classicsId":"arxiv-1910.02054","check":"verified"},{"ref":"arxiv:1909.08053","title":"Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism","year":2019,"url":"https://arxiv.org/abs/1909.08053","classicsId":"arxiv-1909.08053","check":"verified"},{"ref":"arxiv:1811.06965","title":"GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism","year":2018,"url":"https://arxiv.org/abs/1811.06965","classicsId":null,"check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u64-systems/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u64-systems/"}},{"id":"u70-vision","stage":"s7-multimodal-gen","title":{"zh":"视觉 Transformer 与自监督视觉","en":"Vision Transformers & Self-Supervised Vision"},"hours":8,"algorithms":["vit","simclr","mae","jepa","sam","contrastive-learning","unet","faster-rcnn","yolo","detr","swin","dinov2"],"prereqUnits":["u41-transformer","u32-cnn-rnn"],"papers":[{"ref":"arxiv:2010.11929","title":"An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale","year":2020,"url":"https://arxiv.org/abs/2010.11929","classicsId":"arxiv-2010.11929","check":"verified"},{"ref":"arxiv:2002.05709","title":"A Simple Framework for Contrastive Learning of Visual Representations","year":2020,"url":"https://arxiv.org/abs/2002.05709","classicsId":"arxiv-2002.05709","check":"verified"},{"ref":"arxiv:2111.06377","title":"Masked Autoencoders Are Scalable Vision Learners","year":2021,"url":"https://arxiv.org/abs/2111.06377","classicsId":null,"check":"verified"},{"ref":"arxiv:2304.02643","title":"Segment Anything","year":2023,"url":"https://arxiv.org/abs/2304.02643","classicsId":"arxiv-2304.02643","check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u70-vision/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u70-vision/"}},{"id":"u71-vlm","stage":"s7-multimodal-gen","title":{"zh":"CLIP 与视觉语言模型","en":"CLIP & Vision-Language Models"},"hours":8,"algorithms":["clip","siglip","llava","flamingo"],"prereqUnits":["u70-vision","u50-sft-peft"],"papers":[{"ref":"arxiv:2103.00020","title":"Learning Transferable Visual Models From Natural Language Supervision","year":2021,"url":"https://arxiv.org/abs/2103.00020","classicsId":"arxiv-2103.00020","check":"verified"},{"ref":"arxiv:2303.15343","title":"Sigmoid Loss for Language Image Pre-Training","year":2023,"url":"https://arxiv.org/abs/2303.15343","classicsId":null,"check":"verified"},{"ref":"arxiv:2304.08485","title":"Visual Instruction Tuning","year":2023,"url":"https://arxiv.org/abs/2304.08485","classicsId":"arxiv-2304.08485","check":"verified"},{"ref":"arxiv:2204.14198","title":"Flamingo: a Visual Language Model for Few-Shot Learning","year":2022,"url":"https://arxiv.org/abs/2204.14198","classicsId":"arxiv-2204.14198","check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u71-vlm/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u71-vlm/"}},{"id":"u72-diffusion","stage":"s7-multimodal-gen","title":{"zh":"扩散模型、流匹配与图像生成","en":"Diffusion, Flow Matching & Image Generation"},"hours":14,"algorithms":["vae","gan","vq-vae","ddpm","ddim","score-sde","latent-diffusion","cfg","dit","flow-matching","rectified-flow","consistency-models","controlnet","var","normalizing-flows"],"prereqUnits":["u31-training-tricks","u13-probability"],"papers":[{"ref":"arxiv:1312.6114","title":"Auto-Encoding Variational Bayes","year":2013,"url":"https://arxiv.org/abs/1312.6114","classicsId":"arxiv-1312.6114","check":"verified"},{"ref":"arxiv:2006.11239","title":"Denoising Diffusion Probabilistic Models","year":2020,"url":"https://arxiv.org/abs/2006.11239","classicsId":"arxiv-2006.11239","check":"verified"},{"ref":"arxiv:2112.10752","title":"High-Resolution Image Synthesis with Latent Diffusion Models","year":2021,"url":"https://arxiv.org/abs/2112.10752","classicsId":"arxiv-2112.10752","check":"verified"},{"ref":"arxiv:2207.12598","title":"Classifier-Free Diffusion Guidance","year":2022,"url":"https://arxiv.org/abs/2207.12598","classicsId":"arxiv-2207.12598","check":"verified"},{"ref":"arxiv:2212.09748","title":"Scalable Diffusion Models with Transformers","year":2022,"url":"https://arxiv.org/abs/2212.09748","classicsId":"arxiv-2212.09748","check":"verified"},{"ref":"arxiv:2210.02747","title":"Flow Matching for Generative Modeling","year":2022,"url":"https://arxiv.org/abs/2210.02747","classicsId":"arxiv-2210.02747","check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u72-diffusion/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u72-diffusion/"}},{"id":"u73-speech-video-world","stage":"s7-multimodal-gen","title":{"zh":"语音、视频与世界模型","en":"Speech, Video & World Models"},"hours":8,"algorithms":["whisper","dit","nerf-3dgs","world-models","dreamer","genie","jepa","ctc","wav2vec2","valle"],"prereqUnits":["u72-diffusion","u33-rl-basics"],"papers":[{"ref":"arxiv:2212.04356","title":"Robust Speech Recognition via Large-Scale Weak Supervision","year":2022,"url":"https://arxiv.org/abs/2212.04356","classicsId":"arxiv-2212.04356","check":"verified"},{"ref":"arxiv:1803.10122","title":"World Models","year":2018,"url":"https://arxiv.org/abs/1803.10122","classicsId":"arxiv-1803.10122","check":"verified"},{"ref":"arxiv:2301.04104","title":"Mastering Diverse Domains through World Models","year":2023,"url":"https://arxiv.org/abs/2301.04104","classicsId":"arxiv-2301.04104","check":"verified"},{"ref":"arxiv:2402.15391","title":"Genie: Generative Interactive Environments","year":2024,"url":"https://arxiv.org/abs/2402.15391","classicsId":null,"check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u73-speech-video-world/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u73-speech-video-world/"}},{"id":"u80-retrieval","stage":"s8-rag-agents","title":{"zh":"嵌入、向量检索与重排","en":"Embeddings, Vector Search & Reranking"},"hours":8,"algorithms":["bm25","sentence-bert","dpr","colbert","hnsw","product-quantization","faiss","matryoshka","contrastive-learning","cross-encoder-rerank"],"prereqUnits":["u40-tokens","u23-unsupervised"],"papers":[{"ref":"doi:10.1561/1500000019","title":"The Probabilistic Relevance Framework: BM25 and Beyond","year":2009,"url":"https://doi.org/10.1561/1500000019","classicsId":null,"check":"verified"},{"ref":"arxiv:2004.04906","title":"Dense Passage Retrieval for Open-Domain Question Answering","year":2020,"url":"https://arxiv.org/abs/2004.04906","classicsId":null,"check":"verified"},{"ref":"arxiv:1603.09320","title":"Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs","year":2016,"url":"https://arxiv.org/abs/1603.09320","classicsId":null,"check":"verified"},{"ref":"arxiv:2004.12832","title":"ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT","year":2020,"url":"https://arxiv.org/abs/2004.12832","classicsId":null,"check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u80-retrieval/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u80-retrieval/"}},{"id":"u81-rag","stage":"s8-rag-agents","title":{"zh":"RAG 系统工程与评估","en":"RAG Engineering & Evaluation"},"hours":8,"algorithms":["rag","hyde","graphrag","llm-judge"],"prereqUnits":["u80-retrieval"],"papers":[{"ref":"arxiv:2005.11401","title":"Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks","year":2020,"url":"https://arxiv.org/abs/2005.11401","classicsId":"arxiv-2005.11401","check":"verified"},{"ref":"arxiv:2212.10496","title":"Precise Zero-Shot Dense Retrieval without Relevance Labels","year":2022,"url":"https://arxiv.org/abs/2212.10496","classicsId":null,"check":"verified"},{"ref":"arxiv:2404.16130","title":"From Local to Global: A Graph RAG Approach to Query-Focused Summarization","year":2024,"url":"https://arxiv.org/abs/2404.16130","classicsId":null,"check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u81-rag/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u81-rag/"}},{"id":"u82-agents","stage":"s8-rag-agents","title":{"zh":"Agent：工具调用、MCP、规划与记忆","en":"Agents: Tool Use, MCP, Planning & Memory"},"hours":12,"algorithms":["react","toolformer","function-calling","mcp","reflexion","generative-agents","swe-agent","tree-of-thoughts","autogen","memgpt"],"prereqUnits":["u81-rag","u01-prompting"],"papers":[{"ref":"arxiv:2210.03629","title":"ReAct: Synergizing Reasoning and Acting in Language Models","year":2022,"url":"https://arxiv.org/abs/2210.03629","classicsId":"arxiv-2210.03629","check":"verified"},{"ref":"arxiv:2302.04761","title":"Toolformer: Language Models Can Teach Themselves to Use Tools","year":2023,"url":"https://arxiv.org/abs/2302.04761","classicsId":"arxiv-2302.04761","check":"verified"},{"ref":"arxiv:2303.11366","title":"Reflexion: Language Agents with Verbal Reinforcement Learning","year":2023,"url":"https://arxiv.org/abs/2303.11366","classicsId":"arxiv-2303.11366","check":"verified"},{"ref":"arxiv:2405.15793","title":"SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering","year":2024,"url":"https://arxiv.org/abs/2405.15793","classicsId":"arxiv-2405.15793","check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u82-agents/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u82-agents/"}},{"id":"u83-ai-engineering","stage":"s8-rag-agents","title":{"zh":"AI 应用工程：评测、成本与上线","en":"AI Engineering: Evals, Cost & Shipping"},"hours":8,"algorithms":["llm-judge","speculative-decoding","radix-attention"],"prereqUnits":["u82-agents"],"papers":[{"ref":"arxiv:2306.05685","title":"Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena","year":2023,"url":"https://arxiv.org/abs/2306.05685","classicsId":null,"check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u83-ai-engineering/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u83-ai-engineering/"}},{"id":"u90-evaluation","stage":"s9-safety-frontier","title":{"zh":"评测与基准","en":"Evaluation & Benchmarks"},"hours":6,"algorithms":["llm-judge","emergence"],"prereqUnits":["u44-pretraining"],"papers":[{"ref":"arxiv:2306.05685","title":"Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena","year":2023,"url":"https://arxiv.org/abs/2306.05685","classicsId":null,"check":"verified"},{"ref":"arxiv:2206.07682","title":"Emergent Abilities of Large Language Models","year":2022,"url":"https://arxiv.org/abs/2206.07682","classicsId":null,"check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u90-evaluation/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u90-evaluation/"}},{"id":"u91-interpretability","stage":"s9-safety-frontier","title":{"zh":"可解释性：从归因到电路","en":"Interpretability: from Attribution to Circuits"},"hours":10,"algorithms":["attribution","grad-cam","logit-lens","induction-heads","superposition","interp-sae","circuit-tracing","rome","activation-steering"],"prereqUnits":["u41-transformer"],"papers":[{"ref":"arxiv:2209.11895","title":"In-context Learning and Induction Heads","year":2022,"url":"https://arxiv.org/abs/2209.11895","classicsId":null,"check":"verified"},{"ref":"arxiv:2209.10652","title":"Toy Models of Superposition","year":2022,"url":"https://arxiv.org/abs/2209.10652","classicsId":"arxiv-2209.10652","check":"verified"},{"ref":"arxiv:2309.08600","title":"Sparse Autoencoders Find Highly Interpretable Features in Language Models","year":2023,"url":"https://arxiv.org/abs/2309.08600","classicsId":null,"check":"verified"},{"ref":"arxiv:2202.05262","title":"Locating and Editing Factual Associations in GPT","year":2022,"url":"https://arxiv.org/abs/2202.05262","classicsId":null,"check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u91-interpretability/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u91-interpretability/"}},{"id":"u92-security","stage":"s9-safety-frontier","title":{"zh":"安全攻防：越狱、注入与隐私","en":"Security: Jailbreaks, Injection & Privacy"},"hours":6,"algorithms":["adversarial-examples","gcg-jailbreak","prompt-injection","red-teaming","watermarking","dp-sgd","sleeper-agents","membership-inference"],"prereqUnits":["u82-agents"],"papers":[{"ref":"arxiv:1412.6572","title":"Explaining and Harnessing Adversarial Examples","year":2014,"url":"https://arxiv.org/abs/1412.6572","classicsId":"arxiv-1412.6572","check":"verified"},{"ref":"arxiv:2307.15043","title":"Universal and Transferable Adversarial Attacks on Aligned Language Models","year":2023,"url":"https://arxiv.org/abs/2307.15043","classicsId":"arxiv-2307.15043","check":"verified"},{"ref":"arxiv:2302.12173","title":"Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection","year":2023,"url":"https://arxiv.org/abs/2302.12173","classicsId":null,"check":"verified"},{"ref":"arxiv:2301.10226","title":"A Watermark for Large Language Models","year":2023,"url":"https://arxiv.org/abs/2301.10226","classicsId":null,"check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u92-security/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u92-security/"}},{"id":"u93-alignment-governance","stage":"s9-safety-frontier","title":{"zh":"前沿风险、可扩展监督与治理","en":"Frontier Risk, Scalable Oversight & Governance"},"hours":6,"algorithms":["weak-to-strong","debate","constitutional-ai","sleeper-agents"],"prereqUnits":["u51-alignment"],"papers":[{"ref":"arxiv:2312.09390","title":"Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision","year":2023,"url":"https://arxiv.org/abs/2312.09390","classicsId":"arxiv-2312.09390","check":"verified"},{"ref":"arxiv:1805.00899","title":"AI safety via debate","year":2018,"url":"https://arxiv.org/abs/1805.00899","classicsId":null,"check":"verified"},{"ref":"arxiv:2212.08073","title":"Constitutional AI: Harmlessness from AI Feedback","year":2022,"url":"https://arxiv.org/abs/2212.08073","classicsId":"arxiv-2212.08073","check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u93-alignment-governance/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u93-alignment-governance/"}},{"id":"u94-ai4science","stage":"s9-safety-frontier","title":{"zh":"AI for Science 与各学科应用","en":"AI for Science & Domain Applications"},"hours":10,"algorithms":["gnn","alphafold","alphageometry","rlvr"],"prereqUnits":["u41-transformer","u23-unsupervised"],"papers":[{"ref":"arxiv:1609.02907","title":"Semi-Supervised Classification with Graph Convolutional Networks","year":2016,"url":"https://arxiv.org/abs/1609.02907","classicsId":null,"check":"verified"},{"ref":"doi:10.1038/s41586-021-03819-2","title":"Highly accurate protein structure prediction with AlphaFold","year":2021,"url":"https://doi.org/10.1038/s41586-021-03819-2","classicsId":"doi-10-1038-s41586-021-03819-2","check":"verified"},{"ref":"doi:10.1038/s41586-023-06747-5","title":"Solving olympiad geometry without human demonstrations","year":2024,"url":"https://doi.org/10.1038/s41586-023-06747-5","classicsId":"doi-10-1038-s41586-023-06747-5","check":"verified"}],"pages":{"en":"https://aimap.page/en/learn/unit/u94-ai4science/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u94-ai4science/"}},{"id":"u95-frontier","stage":"s9-safety-frontier","title":{"zh":"如何追踪前沿：读论文与复现","en":"Tracking the Frontier: Reading & Reproducing Papers"},"hours":6,"algorithms":["transformer","grpo","flow-matching"],"prereqUnits":["u41-transformer"],"papers":[],"pages":{"en":"https://aimap.page/en/learn/unit/u95-frontier/","zh-hans":"https://aimap.page/zh-hans/learn/unit/u95-frontier/"}}],"tracks":[{"id":"t-everyone","title":{"zh":"零基础使用者","en":"Everyday Users"},"units":["u00-what-is-ai","u01-prompting","u02-hallucination","u40-tokens","u81-rag","u82-agents","u92-security"],"hours":40,"pages":{"en":"https://aimap.page/en/learn/track/t-everyone/","zh-hans":"https://aimap.page/zh-hans/learn/track/t-everyone/"}},{"id":"t-builder","title":{"zh":"应用开发者","en":"Application Developers"},"units":["u00-what-is-ai","u01-prompting","u10-python","u14-pytorch","u40-tokens","u41-transformer","u50-sft-peft","u60-decoding","u62-serving","u63-local","u80-retrieval","u81-rag","u82-agents","u83-ai-engineering","u90-evaluation","u92-security"],"hours":126,"pages":{"en":"https://aimap.page/en/learn/track/t-builder/","zh-hans":"https://aimap.page/zh-hans/learn/track/t-builder/"}},{"id":"t-researcher","title":{"zh":"算法研究者","en":"Algorithm Researchers"},"units":["u10-python","u11-linalg","u12-calculus","u13-probability","u14-pytorch","u20-ml-workflow","u23-unsupervised","u30-nn-backprop","u31-training-tricks","u32-cnn-rnn","u33-rl-basics","u40-tokens","u41-transformer","u42-modern-arch","u43-moe-newarch","u44-pretraining","u50-sft-peft","u51-alignment","u52-reasoning","u61-compression","u64-systems","u72-diffusion","u91-interpretability","u95-frontier"],"hours":257,"pages":{"en":"https://aimap.page/en/learn/track/t-researcher/","zh-hans":"https://aimap.page/zh-hans/learn/track/t-researcher/"}},{"id":"t-leader","title":{"zh":"产品与管理者","en":"Product Managers & Leaders"},"units":["u00-what-is-ai","u01-prompting","u02-hallucination","u44-pretraining","u52-reasoning","u62-serving","u81-rag","u82-agents","u83-ai-engineering","u90-evaluation","u92-security","u93-alignment-governance"],"hours":84,"pages":{"en":"https://aimap.page/en/learn/track/t-leader/","zh-hans":"https://aimap.page/zh-hans/learn/track/t-leader/"}},{"id":"t-scientist","title":{"zh":"各学科研究者","en":"Domain Scientists"},"units":["u01-prompting","u10-python","u13-probability","u14-pytorch","u20-ml-workflow","u21-trees","u23-unsupervised","u30-nn-backprop","u41-transformer","u50-sft-peft","u70-vision","u81-rag","u82-agents","u94-ai4science","u95-frontier"],"hours":143,"pages":{"en":"https://aimap.page/en/learn/track/t-scientist/","zh-hans":"https://aimap.page/zh-hans/learn/track/t-scientist/"}}]}}