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

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)

▲41 ★233 ❝0 2026-10-06

WorldSonus: Bringing Sound to Worlds

WorldSonus adds real-time, spatially aligned stereo sound to world models with mid-stream sound control.

NoizAI · Pengjun Fang, Jingyi Fa, Kam Man Wu +9

Visual generation Multimodal arXiv Code HF

▲32 ★945 ❝0 2026-10-06

Recurrent Looped Transformer

The Recurrent Looped Transformer lets compute depth grow with sequence length, generalising tasks like parity far beyond training length.

Princeton University · Yifan Zhang, Jichen Feng, Shihan Qin +2

LLMs Reasoning arXiv Code HF

▲30 ★2.1k ❝0 2026-10-06

NeMo-DCR: Bit-Exact Delta-Compressed Refit for Scalable Agentic RL at Trillion-Parameter Scale

NeMo-DCR syncs only the ~1% of weights that change per step, bit-exactly, to speed trillion-parameter agentic RL refits.

NVIDIA · Songlin Jiang, Zhiyu Li, Terry Kong +5

Efficient training & inference Reinforcement learning arXiv Code HF

▲30 ★11 ❝1 2026-10-06

Stateless Language Agents: Scaling Long-Horizon Automated Research

Stanford's Stateless Language Agents keep research state outside the agent, tackling duplication and stalls in long-horizon automated research.

Stanford University · Qizheng Zhang, Changxiu Ji, Isaac Sun +15

Agents AI for Science arXiv Code HF

▲27 ★187 ❝1 2026-10-06

On KL-Regularized Policy Optimization

KLPO anchors the KL regulariser at the sampling policy to handle stale, mismatched rollouts in asynchronous LLM-agent RL.

math-ai · Yifan Zhang

Reinforcement learning LLMs arXiv Code HF

▲22 ★9 ❝0 2026-10-06

Consistent Distribution Matching for Data-Free Diffusion Distillation

A data-free, simulation-free diffusion distillation where one student both generates and estimates scores, reaching strong one-step FID.

Vector Institute · Yuxiang Fu, Qi Yan, Zike Wu +4

Visual generation Efficient training & inference arXiv Code HF

▲152 ❝0 2026-10-05

LoGRA: Scaling LLM Reinforcement Learning with Low-Rank Gradient Sketches

LoGRA cuts RL post-training memory with low-rank gradient sketches and keeps training stable with predicted-KL step control.

NVIDIA · Shaokun Zhang, Yifan Zhang, Jian Hu +5

Reinforcement learning Efficient training & inference arXiv HF

▲145 ❝0 2026-10-05

VepAgent: Bridging Causal-Transition via Tool-Augmented Reinforcement Learning for Video Event Prediction

VepAgent combines causal-transition reasoning with tool-augmented RL for video event prediction.

UCSC-VLAA · Qiutong Chen, Yuchan Guo, Zhenlong Yuan +10

Multimodal Reinforcement learning Reasoning arXiv HF

▲119 ★75 ❝0 2026-10-05

From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation

Trace2Env lets a world-model agent role-play an environment reconstructed from historical traces for training and evaluating agents.

Nanyang Technological University Singapore · Quanyu Long, Xiao Chen, Jianda Chen +4

Agents arXiv Code HF

▲46 ★46 ❝0 2026-10-05

MC-Sparse: Deconstructing and Closing the Dense-Sparse Attention Gap in Diffusion Transformers

MC-Sparse is training-free sparse attention that selects individual KV tokens with cached compensation, closing the dense-sparse gap in DiTs.

Tencent Hunyuan · Jiarui Chen, Zeqiang Lai, Jiangshan Wang +5

Visual generation Efficient training & inference arXiv Code HF

▲35 ★49 ❝3 2026-10-05

Towards Looped Models Done Right, Part II: Rethinking at Fixed Points

Driving looped language models towards fixed points enables truncated backprop, KV sharing and cheaper RL with little accuracy loss.

Institute of Foundation Models · Benhao Huang, Chufan Shi, Junlin Chen +4

Efficient training & inference LLMs arXiv Code HF

▲28 ★6 ❝0 2026-10-05

MemPilot: Orchestrating On-Demand Multimodal Memory Curation for LLM Agents

MemPilot trains an RL policy to curate multimodal agent memory on demand under performance–cost–latency preferences.

Nanyang Technological University · Haozhen Zhang, Haodong Yue, Quanyu Long +6

Agents Multimodal arXiv Code HF

▲24 ★4 ❝0 2026-10-05

Optimizing the Optimizer: Language Models Discover Faster Molecular Relaxation Algorithms

Has an LLM agent rewrite molecular geometry optimisers, discovering algorithms needing fewer force calls than the fastest open baseline.

Affiliation unverified · Artem Tsypin, Vladimir Deshchenya, Kuzma Khrabrov +4

AI for Science Agents arXiv Code HF

▲23 ★2.9k ❝0 2026-10-05

RealtimeWAM: One-Step Asynchronous World Action Models

RealtimeWAM generates actions in one step with asynchronous inference to make world-action models real-time.

Nanyang Technological University · Chengtao Lv, Jinyang Du, Shuyi Feng +7

Robotics & embodied AI Efficient training & inference arXiv Code HF

▲174 ★364 ❝0 2026-10-04

Kandinsky 6.0 Video: Foundation Models for Synchronized Video and Audio Generation

Kandinsky 6.0 Video generates 5-second clips with synchronised 44 kHz audio and lip-sync, in 3B and 29B variants.

Kandinsky Lab · Team Kandinsky, Julia Agafonova, Bulat Akhmatov +85

Visual generation Multimodal arXiv Code HF

▲76 ★34 ❝1 2026-10-04

SearchJev: A Fast and Calibrated System-1 Model for Search Agents

SearchJev is a fast, calibrated System 1 model that makes relevance and action decisions for search agents.

Affiliation unverified · Congfeng Cao, Lipeng Zuo, Konstantinos Papakostas +6

Agents Efficient training & inference arXiv Code HF

▲62 ★30 ❝0 2026-10-04

Memadapter: Counterfactual Adaptation Against Memory-induced Sycophancy

Memadapter uses counterfactual adaptation to reduce memory-induced sycophancy, which arises even from correct memories.

DEEP Group at Jilin University · Ruqing Ning, Haibo Meng, Zhishang Xiang +4

Alignment Agents arXiv Code HF

▲55 ★28 ❝0 2026-10-04

EVISKILL: Grounding Skill Evolution in Replayable Evidence

EVISKILL grounds agent skill evolution in replayable evidence to judge which edits should persist.

DEEP Group at Jilin University · Yan Zhou, Yili Wang, Yiwei Dai +2

Agents arXiv Code HF

▲21 ★619 ❝1 2026-10-04

Prism: Dynamic Sparse Attention for Native 2K Joint Video-Audio Generation Model Training

Prism is dynamic sparse attention that focuses on sound-producing regions for training native 2K joint video-audio models.

Tencent Hunyuan · Shuyuan Tu, Qi Tian, Yinming Huang +8

Visual generation Efficient training & inference arXiv Code HF

▲85 ❝0 2026-10-03

Questioning the Questions: Sustaining Self-Evolution in Reasoning Models

Analyses why self-evolving reasoning models collapse (invalid and repeated questions) and proposes question-quality controls.

Huang's INTelligence lab · Jinyuan Li, Chengsong Huang, Langlin Huang +4

Reasoning Reinforcement learning arXiv HF

▲80 ★49 ❝0 2026-10-03

ALoDLM: Adaptively Looped Diffusion Language Models

ALoDLM adds token-adaptive latent recurrence to diffusion language models, narrowing the quality gap with autoregressive models.

Amazon · Liancheng Fang, Zhuowei Li, Youngeun Kim +10

LLMs Efficient training & inference arXiv Code HF

▲26 ❝0 2026-10-03

Super-Resolution in The Right Latent Space: A Frozen Vision-Foundation Substrate

Argues real-world super-resolution works best in a frozen vision-foundation feature space rather than reconstruction latents or pixels.

RTP-LLM · Wanzhou Lei, Cuifeng Shen, Yanjin He +6

Vision & 3D Visual generation arXiv HF

▲105 ❝3 2026-10-02

Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite

Recursive Self-Rewrite finds solutions under specialised harnesses and rewrites them as training trajectories for a general harness.

Tencent Hunyuan · Zongxia Li, Yucheng Shi, Zhongzhi Li +7

Agents Reinforcement learning arXiv HF

▲101 ★56 ❝1 2026-10-02

DuoMatching: Joint-Marginal Distribution Matching for Few-Step Video Generation

DuoMatching matches both joint and per-frame marginal distributions to improve quality and alignment in few-step video generation.

ByteDance · Jiahao Zhan, Yan Wang, Yongrui Ma +5

Visual generation Efficient training & inference arXiv Code HF

▲88 ❝0 2026-10-02

Native Action-Prior Learning from Videos for World Action Models

NAVA-WAM pretrains the action policy directly from observation-only videos, reducing reliance on action-labelled robot data.

Meta · Zhaochong An, Fei Zhang, Menglin Jia +10

Robotics & embodied AI arXiv HF

▲47 ❝0 2026-10-02

Source Preference in the Wild: How LLM Agents Favor Items by Source, and How to Reduce It

Finds 12 agent models systematically favour items from certain sources when shopping, booking or citing, and proposes mitigations.

Seoul National University · Jonghyun Song, Haewon Park, Jeonghoon Shim +2

Alignment Agents arXiv HF

▲34 ★178 ❝4 2026-10-02

4DCodeBench: Benchmarking Agents on Inverse Graphics of Dynamic Scenes

4DCodeBench asks agents to reconstruct dynamic scenes from video as executable graphics programs.

4DCodeBench · Ruihong Shen, Žiga Kovačič, Peter Kulits +6

Benchmarks & data Code & SWE Vision & 3D arXiv Code HF

▲287 ❝0 2026-10-01

RealCompanion: Benchmarking Human Understanding from Reasoning over Longitudinal Real-World Conversations

RealCompanion releases ten real AI-companion relationships (27,218 messages over up to 120 days) to test long-term understanding of a person.

Quis Lab · Arman Behnam, Sunglyoung Kim, Jiayi Yu +2

Benchmarks & data Agents arXiv HF

▲238 ★228 ❝3 2026-10-01

OneStreamer: Unifying Perception, Memory, and Proactive Response in Streaming Video Interaction

OneStreamer lets streaming video LLMs keep hierarchical caption memories and respond proactively once evidence is sufficient.

Nanjing University · Xiangyu Zeng, Yuandong Yang, Zhiqiu Zhang +21

Multimodal arXiv Code HF

▲111 ★30 ❝0 2026-10-01

Sharpening Tax in Post-Training

Finds a 'sharpening tax': pre-trained LLMs with a light harness often beat post-trained ones on agentic coverage despite lower pass@1.

Meta · Changdae Oh, Qi Zeng, Qi Qi +7

Reinforcement learning Agents arXiv Code HF

▲107 ❝1 2026-10-01

SALD: Self-Referenced Advantage Learning for Diffusion Models

SALD lets a diffusion model score each image-caption pair at two noise levels, yielding self-referenced training signals without a teacher.

Indian Institute of Technology, Kanpur · Aryan Das, Surjo Dey, Koushik Biswas +4

Visual generation arXiv HF

▲98 ❝1 2026-10-01

World Action Modeling with Progressive Visual Planning

ProWAM predicts actions plus an ordered sequence of sparse visual subgoals, guiding long-horizon control more efficiently than full video.

Meta · Fei Zhang, Zhaochong An, Duncan Frost +5

Robotics & embodied AI Efficient training & inference arXiv HF

▲92 ★62 ❝0 2026-10-01

Hierarchical Continuous Diffusion Language Models

Hierarchical continuous diffusion LMs tie the denoised continuous state to valid tokens, improving coherence of parallel decoding.

University of Illinois at Urbana-Champaign · Hui Ren, Zihan Li, Chang Liu +2

LLMs arXiv Code HF

▲92 ★32 ❝1 2026-10-01

Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States

PoS has agents maintain explicit belief states of the world and open requirements, avoiding belief traps in long tasks.

alibaba · Yu Luo, Jiamin Jiang, Yimin Zuo +9

Agents arXiv Code HF

▲89 ★34 ❝0 2026-10-01

World Observer: Joint Actor-Observer Generation for Persistent World Modeling

World Observer jointly generates panoramic observer views alongside the actor view so off-screen objects keep evolving consistently.

KAIST AI · Hyunwook Choi, Dahyun Chung, Hyunsung Kim +4

Visual generation arXiv Code HF

▲86 ❝0 2026-10-01

ActiveSaddler: Automated Curriculum Learning for Agent Harness Optimization

Microsoft's ActiveSaddler adapts the training curriculum alongside the evolving agent harness for automated harness optimisation.

Microsoft · Sungho Park, Wonjoong Kim, Jue Zhang +8

Agents arXiv HF

▲83 ★131 ❝3 2026-10-01

UniWAM: Unified World-Action Model

UniWAM unifies a physical reasoner, a world generator and an action predictor in one model for robot control.

HKUSTGZ · Wenxuan Song, Jiayi Chen, Jingbo Wang +13

Robotics & embodied AI Multimodal arXiv Code HF

▲74 ★49 ❝3 2026-10-01

PROWBench: Do Video Models Render What the Program Specifies?

PROWBench tests whether video models faithfully render program-specified world events, aimed at next-gen game engines.

Alaya Lab · Zheng-Hui Huang, Guixu Lin, Yu-Ju Tsai +6

Visual generation Benchmarks & data arXiv Code HF

▲72 ★104 ❝0 2026-10-01

Rethinking World-Action Model for Compositional and In-Context Robotic Manipulation

ViGAR splits long-horizon manipulation into a visual subgoal planner and a subgoal executor for compositional, in-context robot tasks.

DAGroup-PKU · Shukai Gong, Xuanran Zhai, Yintianrun Zhang +16

Robotics & embodied AI Reasoning arXiv Code HF

▲63 ★11 ❝0 2026-10-01

AutoGUIWorld: Image Generators as Visual World Models for GUI Agent

AutoGUIWorld uses image generators as visual world models of GUIs to synthesise diverse interaction trajectories.

Tencent Hunyuan · Cheng Yang, Yifan Wu, Yutao Huang +18

Agents Visual generation arXiv Code HF

▲63 ★75 ❝0 2026-10-01

VeriHarness: Scaling Agentic Verification for Long-Horizon Tasks

VeriHarness turns the base model into a verification agent that uses disagreement across samples to decide which claims to trust.

Google · Caiqi Zhang, Rujun Han, Zifeng Wang +4

Agents Reasoning arXiv Code HF

▲60 ★24 ❝0 2026-10-01

Spatial Memory Intelligence: Endowing World Models with Understanding-Driven Long-Term Memory

Uses MLLM spatial understanding to manage long-term memory so world models stay spatially consistent in long video generation.

Affiliation unverified · Ying Yang, Guiyu Zhang, Lianghua Huang +5

Visual generation Multimodal arXiv Code HF

▲60 ★26 ❝0 2026-10-01

EditHero: A Benchmark for Long-Horizon Part-Level 3D Editing and Vibe Modeling

EditHero is a benchmark for long-horizon, part-level 3D editing, checking that each revision changes only what was asked.

Affiliation unverified · Ruihan Yu, Yu-Ju Tsai, Muyao Niu +9

Vision & 3D Benchmarks & data arXiv Code HF

▲56 ★21 ❝0 2026-10-01

SimuVerity: Benchmarking Agents for Engineering-Grade Simulink Model Generation

SimuVerity's 101 tasks check whether agent-generated Simulink models meet engineering requirements, not just whether they run.

Xi'an Jiaotong University · Ruiqi Zhang, Jiahao Wang, Mingxuan Li +9

Agents Benchmarks & data Code & SWE arXiv Code HF

▲52 ★51 ❝5 2026-10-01

Fewer Tokens, Better Action: GPT-6 Astra Robot Agents with 14% Higher Success Rate but 65% Fewer Tokens

Has robot agents compose primitives into locally retried Python cells, raising success while cutting token use.

DAGroup-PKU · Ruiyang Si, Jianxin Bi, Shunyu Yang +12

Robotics & embodied AI Agents Efficient training & inference arXiv Code HF

▲47 ★164 ❝0 2026-10-01

Science Utopia? Closed-Loop LLM Simulation of Academic Research Ecosystems

SciUtopia simulates academic ecosystems with LLM agents: topic choice, collaboration, review, citation and funding.

Georgia Institute of Technology · Yiqiao Jin, Yiyang Wang, Lucheng Fu +8

Agents AI for Science arXiv Code HF

▲39 ★28 ❝1 2026-10-01

4Director: Controlling Video World Models with Rigid 3D Geometry

Stability AI's 4Director controls camera and object motion in a video world model via explicit rigid meshes and per-frame transforms.

Stability AI · Wei Cao, Hao Zhang, Vikram Voleti +5

Visual generation arXiv Code HF

▲38 ★10 ❝0 2026-10-01

Selection-Based Structured Reasoning: Toward Efficient Multimodal Search Agents

SSR has small models select from predefined reasoning candidates instead of free generation, cutting cost for multimodal search agents.

Carnegie Mellon University · Feiyu Gavin Zhu, Xiaoyu Zhu, Jiqi Yang +9

Reasoning Efficient training & inference Agents arXiv Code HF