ARCHIVE · 2026-10 · 198 PAPERS · PAGE 2/4

Papers of 2026-10

Papers first published on arXiv in 2026-10, newest first. To filter by topic, discipline or organisation, use the archive search.

Agents 83Visual generation 50Efficient training & inference 45Benchmarks & data 40Reinforcement learning 31Multimodal 28Robotics & embodied AI 26LLMs 20Reasoning 18Code & SWE 15Vision & 3D 14AI for Science 9Safety & interpretability 7Alignment 2

Top organisations: alibaba (5) · Tencent Hunyuan (5) · NVIDIA (4) · Nanyang Technological University (4) · Qwen (4)

▲110 ★23 ❝0 2026-10-15

Behavioral Compositionality of Sparse Autoencoder Representations

Introduces SAR to measure whether sparse-autoencoder features compose additively in downstream behaviour, aiding interpretability.

OpenSAE · Haotian Ma, Ziye Mei, Frederic Sala

Safety & interpretability arXiv Code HF

▲93 ★21 ❝0 2026-10-15

Speculative Forcing for Efficient Autoregressive Video Generation

Speculative Forcing treats intermediate estimates as drafts and accepts converged tokens early to speed up autoregressive video generation.

NAVER AI Lab · Hyunwook Choi, Seungryong Kim, Dongyoon Han +1

Visual generation Efficient training & inference arXiv Code HF

▲86 ★4 ❝0 2026-10-15

SCOPE: Semantic Bounded Credit Optimization for Prompt Engineering in Text-to-image Generation

SCOPE, a GRPO-based objective, credits semantic spans rather than sub-word tokens when training prompt rewriters for text-to-image.

University of Rochester · Yufei Jiao, Jinfa Huang, Zeyue Xue +5

Reinforcement learning Visual generation arXiv Code HF

▲64 ★22 ❝0 2026-10-15

VLA-$Δ$: Learning Physical State Transition for Vision-Language-Action Models

VLA-Δ explicitly supervises physical state transitions to reduce VLA overfitting to observation-action pairs.

ZJU-OmniAI · Qi Lu, Jiarun Li, Ce Zhang +5

Robotics & embodied AI arXiv Code HF

▲62 ★36 ❝0 2026-10-15

Mosaic: Multi-Session Agentic Reinforcement Learning

Mosaic generalises agentic RL to multi-session trajectories, matching inference-time context compaction and sub-agent forking.

LARK Lab@HKUST (GZ) · Yilong Xu, Minrui Xu, Xinyu Zhou +3

Reinforcement learning Agents arXiv Code HF

▲55 ★6 ❝0 2026-10-15

Lexical Reference Frames: Interpreting LM Residual Streams in Native Semantics

Lexical reference frames read hidden states through the model's own remaining layers, giving far more faithful readings than fixed lenses.

Affiliation unverified · Jian Gu, Hongyu Zhang, Chunyang Chen +1

Safety & interpretability LLMs arXiv Code HF

▲53 ★4 ❝0 2026-10-15

From Research Goals to Scientific Closure: Evaluating Agents Through Evidence-Supported Physics Research

PhysClosureBench uses 50 expert-verified physics tasks to test whether agents can run investigations to well-supported conclusions.

Shanghai AI Laboratory · Yijia Mao, Song Dai, Tao Chen +14

AI for Science Agents Benchmarks & data arXiv Code HF

▲41 ★14 ❝0 2026-10-15

AutoSFT: An Agentic System for Automated Data Research in Supervised Fine-Tuning

Tsinghua's AutoSFT gives an agent a full data action space—mix, prune, repair, synthesise—to automate SFT data research.

Tsinghua NLP Group · Bingxiang He, Pengqi Sun, Shihao Peng +18

Agents LLMs arXiv Code HF

▲34 ★14 ❝0 2026-10-15

Can Visual Generation Expand What Unified Multimodal Models Know?

A controlled probe shows unified multimodal models can transfer knowledge learned through image generation into text answers, though unevenly.

Z.ai · Wenkai Li, Wenyi Hong, Zhuoyi Yang +1

Multimodal arXiv Code HF

▲25 ★826 ❝0 2026-10-15

Intern-Decision: Efficient Multimodal Decision Making with Autoregressive Masked Language Modeling

Intern-Decision (0.8B–4B) returns structured choices with probabilities instead of free text, making agent decisions faster.

Intern Large Models · Tong Zhu, Zhe Xu, Zefeng He +7

Multimodal Agents Efficient training & inference arXiv Code HF

▲602 ★7 ❝0 2026-10-14

LEV: Thinking Fast and Deep with Looped Decision Models

LEV lets decision models think deeper by looping their layers instead of generating reasoning tokens.

Affiliation unverified · Dingjie Song, Zhiling Yan, Yuxuan Zhang +1

Reasoning Efficient training & inference arXiv Code HF

▲490 ★30 ❝0 2026-10-14

BizAgentBench: Unveiling Hidden Challenges in Deploying Long-Horizon Agents for Real-World Business

BizBench, built on real payment procedures, checks whether long-horizon agents violate policies and constraints mid-execution.

ant-international · Enhan Zhao, Yongqi Tong, Xuejun Yu +6

Agents Benchmarks & data arXiv Code HF

▲364 ★312 ❝0 2026-10-14

Any Time, Any Space: A World-State-Centric Framework for Agentic Video Generation

ATAS compiles screenplays into shots backed by scene and event memories so long agentic videos stay spatially and temporally consistent.

Zhejiang University · Weijie Wang, Haomin Zhang, Ye Li +7

Visual generation Agents arXiv Code HF

▲348 ★187 ❝0 2026-10-14

Player-1: Learning to Play from Synthetic Experience

Player-1 post-trains computer-use agents on synthetic experience so they learn to control games beyond demonstrations.

ARC Lab, Tencent · Mingyu Ouyang, Yiran Wang, Siyuan Hu +9

Agents Reinforcement learning arXiv Code HF

▲149 ★38 ❝0 2026-10-14

MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments

MAI-Gym is an open all-in-one GUI agent platform unifying evaluation, data collection and online RL at up to 20,000 parallel environments.

Tongyi-MAI · Panrong Tong, Quyu Kong, Pengxiang Li +12

Agents Reinforcement learning Benchmarks & data arXiv Code HF

▲131 ★523 ❝0 2026-10-14

InfiniSplat v2: Analyzing and Improving Feed-Forward 3DGS from Any Views

InfiniSplat v2 analyses and improves feed-forward 3D Gaussian reconstruction for better rendering under large viewpoint changes.

zju3dv · Yongzhen Hu, Hao Yu, Jiawei Wang +7

Visual generation Vision & 3D arXiv Code HF

▲110 ★113 ❝0 2026-10-14

Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents

Doc2Harness directly benchmarks whether LLMs can engineer a long-horizon coding-agent harness from a development document.

Institute of Computing Technology · Yuyao Ge, Shenghua Liu, Peizhong Gao +8

Agents Benchmarks & data Code & SWE arXiv Code HF

▲92 ★83 ❝0 2026-10-14

More Methods, Better Agents? A Reality Check on Self-Distillation in Agentic Reinforcement Learning

Compares ten agentic-RL self-distillation methods against GRPO under identical conditions to check real gains and costs.

Zhejiang University · Linhao Zhong, Bozhen Fang, Xuehao Wang +7

Agents Reinforcement learning Benchmarks & data arXiv Code HF

▲91 ★4 ❝0 2026-10-14

Recovering Reliable Supervision via Critique-Guided On-Policy Distillation

CriticOPD has the teacher critique wrong student rollouts before rewriting, restoring reliable supervision in on-policy distillation.

Affiliation unverified · Kaixiang Zhao, Tianrun Yu, Shangzhe Li +4

Efficient training & inference Reasoning arXiv Code HF

▲84 ★10 ❝0 2026-10-14

Speculative Sampling in Interactive Video Generation

Speculative Sampling for interactive video generation caches predictions from large windows to get large-chunk quality at small-chunk latency.

university of science and technology of china · Yubo Huang, Yuxuan Zhang, Jiaming Liu +2

Visual generation Efficient training & inference arXiv Code HF

▲82 ★13 ❝0 2026-10-14

LatentLoop: Scaling Looped MoE Models with Asymmetric Compute Allocation

LatentLoop finds loop visits in looped MoE play asymmetric roles and allocates compute asymmetrically for better results.

Affiliation unverified · Zefeng He, Yunkai Chai, Xiaoyang Chen +5

Efficient training & inference arXiv Code HF

▲73 ❝0 2026-10-14

Cortex: Consolidating Context into Parametric Memory

Cortex consolidates gains from context into parametric memory, keeping them after the context is removed.

ByteDance Seed · Jintian Zhang, Feng He, Xinnian Liang +5

LLMs Efficient training & inference arXiv HF

▲64 ★20 ❝0 2026-10-14

ActionOPD: Distilling Decisions, Not Tokens

ActionOPD distils agents at the level of semantic actions rather than tokens to transfer decisions to smaller models.

Tsinghua-NICS-EFC · Huining Yuan, Xiangmin Yi, Zelai Xu +5

Agents Efficient training & inference arXiv Code HF

▲55 ★37 ❝0 2026-10-14

Seeing, Maintaining, and Learning: The Evolution of Multimodal Agent Memory

A survey of multimodal agent memory, from perceiving and maintaining memories to learned memory-management policies.

Fudan University · Minghao Guo, Qingyue Jiao, Xiujin Liu +77

Multimodal Agents arXiv Code HF

▲51 ★81 ❝0 2026-10-14

Intern Lumina U2: A Multi-Codebook Diffusion Large Language Model for Omni-Visual Understanding and Image Generation

Intern Lumina U2 is a multi-codebook diffusion LLM unifying visual understanding and image generation in one space.

Shanghai AI Laboratory · Qian Yu, Zongyang Qiu, Juncheng Yan +40

Multimodal Visual generation arXiv Code HF

▲43 ★48 ❝1 2026-10-14

Orbia: Do Generated Videos Form a Consistent 3D World?

Orbia is a benchmark for 3D consistency of interactive world models, testing whether scenes persist on revisits.

Tencent · Dongyue Lu, Guian Fang, Hur Lim +20

Visual generation Benchmarks & data arXiv Code HF

▲43 ★111 ❝0 2026-10-14

Towards Agentic Robots: Model, Data, Environment, and Harness

A survey of agentic robots organised around model, data, environment and harness.

Affiliation unverified · Chonghan Qin, Dongming Wu, Kaituo Feng +13

Robotics & embodied AI Agents arXiv Code HF

▲37 ★86 ❝0 2026-10-14

DeltaStreamer: Towards Human-Like Agentic Streaming Video Understanding

DeltaStreamer mimics human observation with change-aware memory and query-driven replay for streaming video understanding.

ByteDance · Boyuan Sun, Deshui Miao, Wenzhao Gao +8

Multimodal Agents arXiv Code HF

▲34 ❝0 2026-10-14

Interlock: From Local Correctness to Task Completion in Long-Horizon Agents

Interlock represents tasks as typed completion graphs to synthesize long-horizon agent tasks and rubrics that check end-to-end completion.

LongCat · Linji Hao, Yanji He, Zihan Wang +6

Agents Benchmarks & data arXiv HF

▲25 ★185 ❝0 2026-10-14

BlenderWorld: Benchmarking Visual Agents for 3D Creation from Real-World References

Qwen's BlenderWorld benchmarks visual agents creating and editing 3D content in Blender from real-world references.

Qwen · Xinyu Fang, Yuchi Wang, Weihao Yu +12

Benchmarks & data Agents Multimodal arXiv Code HF

▲375 ★210 ❝1 2026-10-13

VideoJev: System One Model for Video Understanding and Interaction

VideoJev scores time spans and answer options directly, unifying temporal grounding and video QA for fast decisions.

Multimedia Computing Group-Nanjing University · Yuhan Zhu, Xiangyu Zeng, Qingyi Si +15

Multimodal Efficient training & inference arXiv Code HF

▲371 ★115 ❝2 2026-10-13

Jeva: System One Models from First Principles

Jeva is a decision-native System 1 model that scores candidate options directly instead of generating text answers, making it fast and cacheable.

National University of Singapore · Qi Li, Xinchao Wang

LLMs Efficient training & inference arXiv Code HF

▲217 ★297 ❝0 2026-10-13

Self-Evolving Agent Harness

Formalises self-evolving agent harnesses and derives generalisation and retention bounds for the process.

ZGCAGI · Chuyang Wei, Haoxiang Guan, Minghang Zhu +5

Agents arXiv Code HF

▲214 ★234 ❝0 2026-10-13

Agent-Reflex: Teaching Language Agents to Act Reflectively

Agent-Reflex trains agents in three stages to reflect proactively and recover when tool feedback is noisy or users intervene.

NLPIR Lab @ RUC · Guanting Dong, Xinyu Yang, Jiajie Jin +9

Agents Reinforcement learning arXiv Code HF

▲206 ★6 ❝0 2026-10-13

RoTA-Bench: Evaluating 3D Spatial Reasoning through Interactive Projection Alignment

RoTA-Bench evaluates VLMs' 3D spatial reasoning by having them rotate objects to match target silhouettes interactively.

MAC-AutoML · Qingchuan Ma, Yuexiao Ma, Xiang Meng +7

Multimodal Benchmarks & data Reasoning arXiv Code HF

▲155 ★19 ❝0 2026-10-13

Into the Unknown: Benchmarking AI agents on open scientific research problems

Into the Unknown asks agents to write follow-up papers on unpublished directions from US national-lab scientists to test open research ability.

Carnegie Mellon University · Seungone Kim, Jiayi Geng, Young-Jun Lee +25

Agents AI for Science Benchmarks & data arXiv Code HF

▲137 ★114 ❝0 2026-10-13

AesCode: Aesthetic Code Generation with Decoupled Cross-Modal Rewards

AesCode uses a generated image as aesthetic guidance and decoupled cross-modal rewards to produce better-looking slides and posters as code.

Microsoft · Xin Gao, Yifan Yang, Xin Li +13

Code & SWE Visual generation Reinforcement learning arXiv Code HF

▲133 ★105 ❝0 2026-10-13

Urbex: Agentic Spatial Grounding for City-Scale 3D Scenes

Urbex treats city-scale 3D grounding as active search by a VLM agent that zooms and changes views, optimised with RL.

alibaba · Yu Wang, Rui Dai, Wei Chen +5

Agents Vision & 3D Reinforcement learning arXiv Code HF

▲118 ★5 ❝0 2026-10-13

Pack Light, CarryOn: Building Object-Oriented Context for Long-Horizon Agents

CarryOn organises long-horizon agent context around reusable task objects instead of reprocessing history.

McAuley-Lab · Yuchen Lu, Julian McAuley, Lei Zhang

Agents Efficient training & inference arXiv Code HF

▲110 ★6 ❝0 2026-10-13

GreenWAM: A World Foundation Model for Robotics

GreenWAM is a robotics world foundation model using a shared joint action space across 20 robot families, serving as both world model and policy.

Sber Robotics Center · Gennady Sidorov, Valerii Kornilov, Denis Grankin +2

Robotics & embodied AI arXiv Code HF

▲85 ★4 ❝0 2026-10-13

WorkRover: Scaling Long-Horizon Cowork Agent Training Via Recursive Environment Evolution

WorkRover scales RL for long-horizon cowork agents by recursively evolving tasks, work contexts and verifiers.

Nex AGI · Hang He, Linting Wang, Guanming Liu +14

Agents Reinforcement learning arXiv Code HF

▲79 ★3 ❝0 2026-10-13

Many Right Answers, One Training Example: Removing Responder Noise from GRPO for LLM Recommenders

ReMaRk scores recommendation lists against the union of all replies, removing responder noise from GRPO training of LLM recommenders.

Nanyang Technological University · Xin Li, Hao Jiang, Jinghao Guo +2

Reinforcement learning LLMs arXiv Code HF

▲60 ★135 ❝0 2026-10-13

OmniCompass: A Holistic Benchmark for Omni-Modal Video Grounding and Understanding

OmniCompass evaluates video grounding and understanding on separate vision-language and audio-visual tracks.

VCLab · Xuyang Liu, Shihao Wang, Xinyu Wei +5

Multimodal Benchmarks & data arXiv Code HF

▲52 ★103 ❝0 2026-10-13

LensErase: A Foundation Model for Erase Anything

LensErase is a general erase-anything model trained on 100M pairs, removing objects, text, watermarks and their visual effects.

Microsoft · Yunuo Chen, Xiuyu Wu, Qinhong Yang +11

Visual generation arXiv Code HF

▲41 ★123 ❝0 2026-10-13

FlashVSR+: Efficient Video Super-Resolution via Hybrid Attention and Reward-Yield Distribution Matching

FlashVSR+ speeds up one-step video super-resolution with hybrid attention and aligns it to human preference via reward-weighted distribution matching.

MMLab-CUHK · Ruchang Yao, Shi Guo, Junhao Zhuang +7

Visual generation Efficient training & inference arXiv Code HF

▲38 ★386 ❝0 2026-10-13