ARCHIVE · 2026-09 · 113 PAPERS · PAGE 2/3

Papers of 2026-09

Papers first published on arXiv in 2026-09, newest first. To filter by topic, discipline or organisation, use the archive search.

Agents 39Efficient training & inference 24Multimodal 24Benchmarks & data 22Reinforcement learning 18Visual generation 17Robotics & embodied AI 14Reasoning 10AI for Science 10Code & SWE 8LLMs 7Vision & 3D 7Alignment 6Safety & interpretability 4

Top organisations: Zhejiang University (5) · National University of Singapore (5) · Peking University (5) · Carnegie Mellon University (3) · Google (3)

▲140 ★82 ❝0 2026-09-28

What Makes World Action Models Generalize? An Empirical Study of Test-Time Future Modeling

Finds latent world-action models lose the generalisation benefits of explicitly generating future frames at test time.

Tsinghua-LeapLab · Renping Zhou, Zanlin Ni, Zihao Fan +8

Robotics & embodied AI arXiv Code HF

▲129 ❝0 2026-09-28

LEGO-Anything: Coding Agents for 3D Scene Reconstruction

LEGO-Anything has a coding agent iteratively write Blender code to reconstruct editable 3D scenes from a single image.

Amazon Web Services · Xirui Li, Peng Shi, Mingwen Dong +7

Code & SWE Vision & 3D Agents arXiv HF

▲112 ★54 ❝4 2026-09-28

Self-Evolving Coding Agents: From Digital Programs to Physical-World Intelligence

Argues for expressing robot tasks as inspectable, revisable code so self-evolving coding agents can move into the physical world.

HexaFuture · Hongcheng Gao, Jingjing Zhou, Zelin Zheng +10

Robotics & embodied AI Code & SWE Agents arXiv Code HF

▲108 ❝0 2026-09-28

Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression

Finds chunked KV-cache compression causes phase sensitivity: retrieval accuracy varies with position relative to window boundaries.

ByteDance Seed · Xingyu Zhu, Pu, Yi +6

Efficient training & inference arXiv HF

▲87 ★5 ❝1 2026-09-28

Agent Priors-guided Policy Learning

Exposes each skill policy's structural prior to the agent so few-shot robots generalise both compositionally and per skill.

National University of Singapore · Puming Jiang, Tianrun Hu, Haozhe Du +4

Robotics & embodied AI Agents arXiv Code HF

▲86 ★20 ❝0 2026-09-28

LLMs are General Asynchronous Agents

Proposes general asynchronous agents that accept new inputs while thinking or acting, covering voice, embodied and monitoring settings.

Yandex Research · George Yakushev, Denis Mazur, Vladimir Bartenev +4

Agents arXiv Code HF

▲74 ★23 ❝0 2026-09-28

LongCat-DeepResearch Technical Report

Technical report for LongCat-DeepResearch, a multi-agent workflow that plans, researches in parallel and revises evidence-grounded reports.

LongCat · Meituan LongCat Team, He Zhu, Yue Xu +28

Agents arXiv Code HF

▲65 ★12 ❝0 2026-09-28

PDE-JEPA: Predictive Representation Learning of Latent Dynamics Modeling for Parametric PDEs

PDE-JEPA learns latent dynamics of parametric PDEs with predictive (JEPA-style) pretraining instead of reconstruction.

Zhejiang University · Zhentao Tan, Jianrong Zhang, Ruijie Quan +1

AI for Science Efficient training & inference arXiv Code HF

▲62 ★102 ❝1 2026-09-28

Improving Test-Time Scaling with Adaptive Looped Transformers

Studies looped transformers under test-time scaling and proposes adaptive looping to beat baselines at matched compute.

Tsinghua-NICS-EFC · Yichen You, Tianyu Fu, Aosong Feng +4

Reasoning Efficient training & inference arXiv Code HF

▲50 ★31 ❝0 2026-09-28

Learning Native Reflection in Unified Models with Interleaved Reinforcement Learning

Interleaved RL jointly trains reflection text and image generation so unified models can diagnose and repair their own images.

MMLab@NTU · Yijia Fan, Ziqi Huang, Zhongang Cai +5

Multimodal Reinforcement learning Visual generation arXiv Code HF

▲748 ★5.5k ❝1 2026-09-27

Raven: The Harness of Harnesses for Composable Agentic Intelligence

Raven is an open multi-agent ecosystem that automatically builds, evolves and orchestrates specialised agent harnesses across domains.

EverMind · EverMind AI

Agents arXiv Code HF

▲296 ★1.3k ❝0 2026-09-27

VisionHOPE: Visual Backbones as Self-Modifying Learning Systems

VisionHOPE turns a visual backbone into a self-modifying learner whose memory and update rules co-evolve within each image.

Mininglamp Technology · Siran Peng, Tianshuo Zhang, Tianyu Fu +8

Vision & 3D arXiv Code HF

▲223 ★12k ❝1 2026-09-27

YuE2: Unifying Symbolic and Audio Music Generation at Frontier Quality

YuE2 plans a readable score first and then renders full songs, unifying symbolic and audio music generation.

Multimodal Art Projection · Ruibin Yuan, Jiahao Pan, Junyan Jiang +32

Multimodal Visual generation arXiv Code HF

▲72 ★13 ❝0 2026-09-27

TraceDance: An Automated System for Building Agent Behavior Benchmarks from Real-World Agent Deployment Traces

TraceDance automatically builds targeted benchmarks for specific undesirable agent behaviours from real deployment traces.

ByteDance · Dehai Min, Daoan Zhang, Yiming Zeng +13

Agents Benchmarks & data Safety & interpretability arXiv Code HF

▲46 ★72 ❝0 2026-09-27

Relevance Does Not Imply Applicability: Experience Activation for Personal GUI Agents

Finds retrieved history mainly helps a personal GUI agent's first step, and proposes activating experience only when it applies.

Nanyang Technological University · Fuyao Zhang, Xuan Wang, Zherui Li +3

Agents arXiv Code HF

▲35 ★37 ❝0 2026-09-27

Diffusion Reward Models

Tsinghua's Diffusion Reward Model recasts reward modelling as conditional density estimation to capture multimodal human preferences.

Tsinghua NLP Group · Xiangyang Wang, Bingxiang He, Zeyuan Liu +13

Alignment Reinforcement learning arXiv Code HF

▲327 ❝0 2026-09-26

Scaling Properties of Same-Family On-Policy Distillation

Studies scaling of same-family on-policy distillation, finding accuracy rises roughly linearly with KL distance early in training.

Zhejiang University · Yuntai Bao, Qinfeng Li, Guoqing Jiang +5

Efficient training & inference Reinforcement learning arXiv HF

▲134 ★7 ❝0 2026-09-26

VoxMem: Benchmarking Multimodal Memory in Large Audio Language Models

VoxMem benchmarks whether audio LLMs remember who said what, how, and what was audible—information beyond the transcript.

The University of Melbourne · Yang Xiao, Vidhyasaharan Sethu, Eun-Jung Holden +1

Multimodal Benchmarks & data arXiv Code HF

▲103 ★7 ❝0 2026-09-26

Prefill-Free Cross-Family KV Cache Transfer for Heterogeneous Multi-Agent LLMs

HeteroFold transfers KV caches across different model families, avoiding repeated prefill in heterogeneous multi-agent systems.

University of Southern California · Vincent-Daniel Yun, Woosang Lim, Haneul Yoo +3

Efficient training & inference Agents arXiv Code HF

▲95 ❝1 2026-09-26

Groupwise Agentic Grading and Advantage Redistribution for Code Agent RL

GAGAR redistributes advantages by implementation quality in code-agent RL, rewarding cleaner, targeted changes.

Xiaomi MiMo · Jinhao Dong, Liang Zhao, Zihao Yue +7

Code & SWE Reinforcement learning arXiv HF

▲81 ★30 ❝0 2026-09-26

DoSA: Dots Sparse Attention

DoSA uses bounds to reject keys early, speeding up sparse attention without changing exact Top-K selection.

dots studio · Fanxu Meng, Taosong Fang

Efficient training & inference arXiv Code HF

▲80 ★2 ❝1 2026-09-26

X-Tree: Tokenizing Reusable Experience for Efficient Agent Generalization

X-Tree tokenises recurring sub-procedures in agent trajectories and trains on that hierarchy, improving generalisation without LLM calls.

University of Waterloo · Sitao Cheng, Xunjian Yin, Zhiyuan Sun +4

Agents Efficient training & inference arXiv Code HF

▲75 ★763 ❝0 2026-09-26

MassAlloc Attention: Let Attention Allocate Its Own Compute

MALA attention allocates post-score computation by normalised contribution, skipping negligible work to save compute.

Beijing Academy of Artificial Intelligence · Jingze Shi, Zhangyang Peng, Xianduo Li +6

Efficient training & inference arXiv Code HF

▲67 ★763 ❝0 2026-09-26

CoWindow Attention: Full Causal Coverage Is a Collective Property

CoWA attention splits long-range history across KV heads so each is sparse yet together they cover the full causal context.

Beijing Academy of Artificial Intelligence · Jingze Shi, Zhangyang Peng, Xianduo Li +6

Efficient training & inference arXiv Code HF

▲485 ★200 ❝0 2026-09-25

Omni-IO Skills: Harnessing Your Agent Omni-Native

Omni-IO Skills makes existing agents omni-modal through hierarchical skills, a common execution interface and an asset registry.

National University of Singapore · Yanlin Li, Mingyang Hao, Shengqiong Wu +3

Agents Multimodal arXiv Code HF

▲138 ★17 ❝1 2026-09-25

FuseReg: Regularizing Layer Fusion Mitigates the Reconstruction-Generation Gap in Representation Autoencoders

FuseReg learns regularised layer fusion to ease the reconstruction-generation trade-off in representation autoencoders.

USC Physical Superintelligence (PSI) Lab · Hongyang Du, Yunfei Xie, Junjie Ye +13

Visual generation arXiv Code HF

▲90 ❝1 2026-09-25

Duplex-MPE: Benchmarking Multi-Party Interaction in Full-Duplex Dialogue

Duplex-MPE uses 2,000 multi-party scenarios to test when a full-duplex voice assistant should answer, stay silent or stop.

Peking University · Chengqian Ma, Wenhao Feng, Weixuan Jin +7

Multimodal Benchmarks & data arXiv HF

▲66 ★5 ❝1 2026-09-25

Decentralized Master-Mind: Joint Action Refinement through Iterative Intent Denoising in Multi-Agent Pathfinding

DMM refines joint actions in decentralised multi-agent pathfinding by iterative intent denoising, avoiding incompatible combinations.

Affiliation unverified · Valeriy Vyaltsev, Anton Andreychuk, Taisia Zlotnikova +3

Robotics & embodied AI Reinforcement learning arXiv Code HF

▲273 ★18 ❝0 2026-09-24

Post-Training Leaves Behavioral Shadows on Unrelated Decisions

Shows post-training leaves behavioural shadows in unrelated outputs; a single teacher word per prompt can transfer capabilities to a student.

Peking University · Ziyang Zhang, Yubin Jing, Yuanhao Zeng +3

LLMs Safety & interpretability arXiv Code HF

▲34 ★111 ❝1 2026-09-24

Qwen-Planner-Agent: A Closed-Loop AI-for-AI Framework for Real-World Mobile Planner Agents

Qwen-Planner-Agent uses a closed-loop AI-for-AI framework to iteratively build real-world mobile planner agents.

Tongyi-MAI · Tingyu Qu, Weigao Sun, Yuecheng Liu +23

Agents Reinforcement learning arXiv Code HF

▲31 ★26 ❝1 2026-09-24

ExplorationBench: Measuring AI Systems' Exploration in Verifiable Alien Worlds

Tencent Hunyuan's ExplorationBench measures scientific exploration in verifiable alien worlds, ruling out recall from pre-training.

Tencent Hunyuan · Ming Zhang, Zhenghao Xiang, Peizhong Gao +17

AI for Science Benchmarks & data Agents arXiv Code HF

▲29 ★28 ❝12 2026-09-24

Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures

Tests whether Jev, a calibrated decision model, can zero-shot detect ten alignment failures such as sycophancy, jailbreaks and deception.

Affiliation unverified · Ruoqi Guo, Yi Liu, Gelei Deng +6

Safety & interpretability Alignment arXiv Code HF

▲23 ❝0 2026-09-24

Where Does Retrieval-Based Open-Ended Evaluation Fail? Automatic Taxonomy Induction from Long-Form Medical Answer Factuality Verification

Builds taxonomies of where retrieval-based factuality checking of long medical answers fails.

Center for Language and Speech Processing @ JHU · Heyuan Huang, Jirui Dai, Alexandra DeLucia +5

AI for Science Benchmarks & data arXiv HF

▲21 ★9 ❝0 2026-09-24

IndicBankBench: Evaluating Safety and Reliability of Language Model Assistants in Indian Retail Banking

IndicBankBench evaluates Indian retail-banking assistants on 799 cases across safety, tool use and answer quality.

National Payments Corporation of India · Suvradip Paul, Chandra Bhushan, Harsh Sharma +4

Safety & interpretability Benchmarks & data arXiv Code HF

▲167 ★97 ❝0 2026-09-23

Training Object Permanence in World Models

Builds 150 cognitive-science tasks to test and train object permanence in video world models.

Carnegie Mellon University · Haotian Zhang, Fengyuan Yu, Dezhi Luo +28

Visual generation Reasoning arXiv Code HF

▲66 ★89 ❝0 2026-09-23

The Past Frames the Future: Memory for Autoregressive Video Generation

Examines memory for autoregressive video generation: keeping identities, states and causal changes within bounded context.

Affiliation unverified · Harold Haodong Chen, Rongjin Guo, Disen Lan +22

Visual generation arXiv Code HF

▲163 ★36 ❝1 2026-09-22

The Tasteful Agent: Measuring and Improving Taste in Long-Horizon Tasks

Taste-Bench measures an agent's 'taste' for long-horizon decisions and explores ways to improve it.

Microsoft · Wenbo Pan, Zhichao Liu, Shujie Liu +6

Agents Benchmarks & data arXiv Code HF

▲103 ★114 ❝0 2026-09-22

SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue

SpeakerMem-R1 builds speaker-centred dual-track memory for multi-party dialogue, tracking who said what about whom.

Zhejiang University · Haobo Zheng, Tan Tang, Yan Chen +2

Agents Reinforcement learning arXiv Code HF

▲92 ★28 ❝0 2026-09-22

Disaggregated Quantization: Specializing LLM Prefill and Decode

Disaggregated quantisation specialises formats and weights for prefill and decode separately to balance speed and accuracy.

IST Austria Distributed Algorithms and Systems Lab · Andrei Panferov, Maximilian Kleinegger, Sweta Priyadarshi +2

Efficient training & inference arXiv Code HF

▲44 ❝40 2026-09-22

JEV-as-a-Judge: Accept When Confident, Escalate When Unsure

Uses a decision-only judge as a cheap first pass and escalates to stronger judges only when unsure.

Carnegie Mellon University · Yubo Li, Yidi Miao, Ramayya Krishnan +1

Benchmarks & data Efficient training & inference arXiv HF

▲220 ★1.6k ❝22 2026-09-21

RRSI: Regularized Recursive Self-Improvement of Agent Harnesses

Adds regularisation to recursive self-improvement of agent harnesses so gains survive out-of-distribution tests.

Google · Peng Xia, Rujun Han, Zifeng Wang +11

Agents Benchmarks & data arXiv Code HF

▲161 ★512 ❝1 2026-09-21

WorldCrafter: Consistent Video World Model with Implicit 3D-aware Memory

WorldCrafter adds a camera-queryable implicit 3D memory to a video world model to keep long explorations consistent.

ARC Lab, Tencent · Wangbo Yu, Kunhao Liu, Wenbo Hu +8

Visual generation arXiv Code HF

▲104 ❝1 2026-09-21

RULER: Instance-aware Rubric Rewards for SVG Generation

RULER uses instance-aware multi-axis rubrics as rewards for SVG generation, reducing the reward hacking seen with scalar metrics.

inclusionAI · Hangyu Ran, Yuhao Zheng, Yingying Zhang +2

Reinforcement learning Visual generation arXiv HF

▲78 ★526 ❝0 2026-09-21

GAE: Learning a Geometry-Native Latent Space for 3D-Consistent World Generation

GAE reparameterises geometry foundation-model features into a compact latent space so generated worlds stay 3D-consistent.

ARC Lab, Tencent · Jiahao Lu, Minghao Yin, Wenbo Hu +5

Visual generation Vision & 3D arXiv Code HF

▲77 ★58 ❝0 2026-09-21

Ovis-Embedding: Pushing the Frontiers of Universal Omni-Modal Embeddings

Ovis-Embedding encodes text, images, video and audio into one shared embedding space using an omni-modal backbone.

ATH-MaaS · Ovis-Embedding Team

Multimodal arXiv Code HF

▲73 ★572 ❝0 2026-09-21

GameHorizon Suite: Multi-Horizon Data and Evaluation in Gameplay

GameHorizon provides multi-horizon gameplay data and evaluation to measure planning and control across model families.

Tencent · Yiran Wang, Xingyilang Yin, Junfu Pu +12

Benchmarks & data Agents arXiv Code HF

▲73 ❝0 2026-09-21

VideoGen-Agent: Reinforcing Video Generation Agents

VideoGen-Agent trains a multimodal agent with multitask agentic RL to use tools for knowledge- and identity-heavy video prompts.

Princeton University · Binxu Li, Haoyi Duan, Yuhui Zhang +9

Visual generation Agents Reinforcement learning arXiv HF

▲56 ★12 ❝2 2026-09-21

All-in-One Multilingual Scene Text Recognition with Script-aware Mixture-of-Experts

A script-aware mixture-of-experts gives one multilingual scene-text recogniser that is simpler than per-language models and lighter than VLMs.

Fudan University · Xingsong Ye, Yongkun Du, Jiaxin Zhang +6

Vision & 3D Efficient training & inference arXiv Code HF

▲55 ★63 ❝0 2026-09-21

onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents via Token-Level Correction

onPanda is an annotation tool where annotators fix the first bad token and let the model continue, cheaply producing alignment data.

StepFun · Lei Yang, Mengyin Liu, Jia Wang +9

Alignment Agents arXiv Code HF

▲38 ★115 ❝6 2026-09-21

Harness-Zero: Harness Distillation via Agent-as-Harness

Harness-Zero distils the gains of specialised agent harnesses into model weights so they persist under a single fixed harness.

Peking University · Haoran Ye, Yuxing Lu, Haonan Dong +2

Agents Efficient training & inference arXiv Code HF