#76 ▲63 ★210 ❝3 2026-09-30
Video Generation Models: A Survey of Post-Training and Alignment
视频生成模型后训练与对齐的综述，梳理时序一致性、物理和安全约束等难点。
Arizona State University · Chaoyu Li, Xiaoyi Gu, Yogesh Kulkarni +10
TOPIC · ALIGNMENT · 8 PAPERS
对齐
大语言模型推理对齐智能体多模态视觉生成视觉理解与 3D机器人与具身高效训练与推理安全与可解释AI for Science代码与软件工程评测与数据强化学习
近 30 天 8
#124 ▲104 ★35 ❝0 2026-09-29
Think Before You Score: Thinking Reward Model for Visual Generation
思考型奖励模型先为每个示例制定评分要点再打分，给视觉生成提供更细的奖励信号。
机构待核实 · Xuehai Bai, Zhenchen Tang, Yang Shi +9
#125 ▲55 ★63 ❝0 2026-09-21
onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents via Token-Level Correction
onPanda 是对齐数据标注工具：找到第一个错 token 改掉后让模型续写，低成本标注 Agent 轨迹。
StepFun · Lei Yang, Mengyin Liu, Jia Wang +9
#160 ▲68 ★22 ❝0 2026-09-30
Scoring Higher, Answering Worse: Mitigating Reward Hacking in Rubric-Based RL via Protocol-Level Rubrics
指出基于评分细则的 RL 用加权求和会被“刷分”（以临床问诊为例），并提出协议级细则加以缓解。
ByteDance · Maoqi Liu, Junwei He, Bowen Zhang +5
#170 ▲29 ★28 ❝12 2026-09-24
Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures
测试经 RL 校准的决策模型 Jev 能否零样本检测谄媚、越狱、欺骗等十类对齐失败。
机构待核实 · Ruoqi Guo, Yi Liu, Gelei Deng +6
#216 ▲62 ★30 ❝0 2026-10-04
Memadapter: Counterfactual Adaptation Against Memory-induced Sycophancy
Memadapter 用反事实适配缓解长期记忆引发的谄媚，即使记忆本身正确也会让 Agent 迎合用户。
DEEP Group at Jilin University · Ruqing Ning, Haibo Meng, Zhishang Xiang +4
#256 ▲35 ★37 ❝0 2026-09-27
Diffusion Reward Models
清华 DRM 把奖励建模改成条件密度估计，用扩散模型刻画人类偏好的多峰分布。
Tsinghua NLP Group · Xiangyang Wang, Bingxiang He, Zeyuan Liu +13
#292 ▲47 ❝0 2026-10-02
Source Preference in the Wild: How LLM Agents Favor Items by Source, and How to Reduce It
发现 12 个 Agent 模型在购物、订酒店、引文献时都会偏好某些来源，并提出缓解方法。
Seoul National University · Jonghyun Song, Haewon Park, Jeonghoon Shim +2