ARCHIVE · 2026-09 · 113 PAPERS · PAGE 1/3

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)

▲642 ❝2 2026-09-30

False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents

Identifies co-cheating in self-evolving search agents, where proposer and solver converge on shared errors, and mitigates it with multi-sample verification.

Rutgers University · Meijia Chen, Hao Li, Zheng Lu +12

Agents Reinforcement learning arXiv HF

▲310 ★140 ❝2 2026-09-30

GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives

GroundingPI is a 4B grounding foundation model that outputs points and boxes as tokens to give robots fast, precise targeting.

XPENG AI · Qize Yu, Lianrui Fan, Boyu Chen +23

Robotics & embodied AI Vision & 3D arXiv Code HF

▲266 ❝0 2026-09-30

UniEvo-VL: An On-policy Self-Distillation Training Recipe for Multimodal Model Self-improvement

UniEvo-VL has one multimodal model act as both teacher and student, using its own critiques for on-policy self-distillation.

Stanford NLP · Fang Wu, Da Xing, Yanjie Huang +16

Multimodal arXiv HF

▲124 ★30 ❝0 2026-09-30

FrameMorrow: Future-guided Frame Selection with Prospective Tokens for Long-Horizon Video Generation

FrameMorrow selects history frames by future relevance via prospective tokens for efficient, consistent long video generation.

National University of Singapore · Bo Yin, Xiaobin Hu, Jiaqi Zhao +1

Visual generation Efficient training & inference arXiv Code HF

▲123 ❝0 2026-09-30

GraphForge: Training Working Agents with Graph-Anchored Workspace Synthesis

GraphForge synthesises verifiable workspace tasks grounded in real files via evidence graphs to train working agents.

University of Science and Technology of China · Qisheng Su, Hanchen Wang, Guanru Zhu +9

Agents Benchmarks & data arXiv HF

▲120 ❝1 2026-09-30

Mid-Harness: Scaling Actions Between Model and Harness for Terminal Agents

Mid-Harness samples and verifies candidate actions before execution, improving terminal-agent reliability.

NVIDIA · Minki Kang, Ryo Hachiuma, Shaokun Zhang +8

Agents Efficient training & inference arXiv HF

▲112 ★7 ❝1 2026-09-30

EvoDuet: Bilevel Co-Evolution of Web Searching and Task Solving for Scientific Discovery

EvoDuet co-evolves solutions and web-search queries in a bilevel loop so LLM evolutionary search doesn't stall for lack of knowledge.

Minnesota NLP · Young-Jun Lee, Jinheon Baek, Soyeong Jeong +5

AI for Science Agents arXiv Code HF

▲106 ★4 ❝0 2026-09-30

WorldAuditBench: Interactive 3D World Auditing with Multimodal Agents

WorldAuditBench has multimodal agents navigate 3D worlds to find anomalies like floating objects or walk-through walls.

University of California, Santa Barbara · Ziyan Jiang, Jingbo Yang, Jiabao Ji +5

Multimodal Agents Benchmarks & data arXiv Code HF

▲97 ★146 ❝0 2026-09-30

RSIGame: Autonomous Agentic Game Development with Recursive Self-improvement

RSIGame improves agent-built games with local and global recursive loops to avoid overfitting a few test cases.

RSIGame · Wenyi Wu, Minghao Fu, Jieyu You +10

Agents Code & SWE arXiv Code HF

▲94 ★78 ❝4 2026-09-30

GroundAnything: Reconciling Parallel Decoding with Precise Visual Grounding at Flash Speed

GroundAnything decodes grounding outputs in parallel with bidirectional diffusion, keeping precision while being much faster.

XPENG AI · Qize Yu, Lianrui Fan, Bowen Ping +19

Vision & 3D Efficient training & inference arXiv Code HF

▲75 ★11 ❝0 2026-09-30

EgoTools: Towards Tool-Centric Reasoning in Real-World Egocentric Videos

EgoTools targets tool-centric reasoning in real egocentric videos, covering hand-tool-object relations and procedures.

Ropedia · Shulin Tian, Junsu Kim, Shuai Liu +17

Multimodal Reasoning Benchmarks & data arXiv Code HF

▲69 ❝1 2026-09-30

PhysVista: Benchmarking Physical Intelligence in VLMs via a Perception-Reasoning-Assessment Loop

PhysVista evaluates VLMs' physical intelligence through a perception-reasoning-assessment loop.

university of science and technology of china · Xinge Peng, Yiting Lu, Tianwu Zhi +4

Multimodal Benchmarks & data Reasoning arXiv HF

▲68 ★22 ❝0 2026-09-30

Scoring Higher, Answering Worse: Mitigating Reward Hacking in Rubric-Based RL via Protocol-Level Rubrics

Shows weighted-sum rubric rewards can be hacked (e.g. in clinical consultation) and proposes protocol-level rubrics to fix it.

ByteDance · Maoqi Liu, Junwei He, Bowen Zhang +5

Reinforcement learning Alignment AI for Science arXiv Code HF

▲68 ★14 ❝0 2026-09-30

Hear the World in Stereo: Learning Dynamic Spatial Correspondence for Immersive Joint Video-Audio Generation

StereoBind ties stereo sound position to on-screen source motion for immersive joint video-audio generation.

Affiliation unverified · Hanmo Chen, Chengcheng Liu, Tianxiao Chen +7

Multimodal Visual generation arXiv Code HF

▲65 ★5 ❝3 2026-09-30

More Choices, Fewer Decisions: Ordinal-Scale Bias in JEV-like Direct-Decision Models

Finds JEV-like direct-decision models under-use ordinal scales, over-predicting middle labels.

Jilin University · Tianxiang Gao, Jinzhe Li, Zhiyuan Li +2

Benchmarks & data arXiv Code HF

▲63 ★210 ❝3 2026-09-30

Video Generation Models: A Survey of Post-Training and Alignment

A survey of post-training and alignment for video generation, covering temporal coherence, physics and safety challenges.

Arizona State University · Chaoyu Li, Xiaoyi Gu, Yogesh Kulkarni +10

Visual generation Alignment arXiv Code HF

▲63 ★14 ❝1 2026-09-30

Science or Slop?: Benchmarking and Mitigating Scientific Slop in AI-Generated Papers

SciSlopBench collects 390 AI-generated papers to detect plausible-looking but broken 'scientific slop' and explores mitigations.

Seoul National University · Yerim Oh, Young-Jun Lee, Jaewoo Ahn +2

Benchmarks & data AI for Science arXiv Code HF

▲51 ★112 ❝0 2026-09-30

Does Learning Protein Folding Generalize to Broader Reasoning?

Asks whether learning protein folding teaches LLMs reusable spatial and structural reasoning.

Shanghai JiaoTong University · Yong Liu, Zhanpeng Shi, Yizhou Dang +4

AI for Science Reasoning arXiv Code HF

▲511 ★7 ❝2 2026-09-29

The Teacher Is a Direction, Not a Destination: Extrapolating RL-Induced Representation Residuals in On-Policy Distillation

Extrapolates RL-induced hidden-representation residuals in on-policy distillation so students can surpass teachers more stably.

Affiliation unverified · Hao Li, MeiJia Chen, Weijie Ren +4

Efficient training & inference Reinforcement learning arXiv Code HF

▲417 ★261 ❝0 2026-09-29

LoopVL: Recurrent Visual Intelligence

LoopVL brings looped (recurrent) Transformers to vision-language models, iteratively refining a shared vision-language state.

Renmin University of China · Zhe Qian, Ziyang Gong, Zhongxing Xu +13

Multimodal Reasoning arXiv Code HF

▲140 ❝0 2026-09-29

Adaptive Reward Routing: Dynamic Multi-Reward Optimization for Joint Audio-Video Diffusion via Forward-Process RL

Adaptive Reward Routing dynamically adjusts where updates act and how rewards combine to improve joint audio-video diffusion.

Tencent · Songlin Yang, Xiaotong Zhao, Jiacheng Zhang +5

Visual generation Reinforcement learning arXiv HF

▲134 ★39 ❝0 2026-09-29

AREX-2: Advancing Self-Improving Agents through Long-Horizon Reflective Tasks

AREX-2 trains on long improvement trajectories from ML and algorithm tasks so agents iteratively refine solutions at test time.

Beijing Academy of Artificial Intelligence · Hongjin Qian, Chaofan Li, Kun Luo +11

Agents Reinforcement learning arXiv Code HF

▲131 ★101 ❝0 2026-09-29

STEPQuant: When and Where Errors Matter in Delta-Rule Recurrent State Quantization

STEPQuant studies where and when quantisation errors matter in linear-attention recurrent states, enabling low-precision state storage.

Zhejiang University · Bingchen Yao, Haobo Xu, Haokun Lin +6

Efficient training & inference arXiv Code HF

▲116 ★59 ❝0 2026-09-29

Beyond the Timeline: Augmenting Long-Video Memory with Grounded Entity Biographies

GEB groups observations of the same physical object into entity biographies to answer questions over long videos.

Amazon Science · Hui Ren, Lei Fan, Henry Pao +5

Multimodal arXiv Code HF

▲115 ❝0 2026-09-29

MotorMind: Scaffolding General Vision Language Models for Zero-Shot Robot Manipulation

MotorMind scaffolds general vision-language models for zero-shot robot manipulation without specialised VLA training.

University of Illinois at Urbana-Champaign · Bingxuan Li, Siqi Song, Yizhuo Wu +3

Robotics & embodied AI Multimodal arXiv HF

▲104 ★35 ❝0 2026-09-29

Think Before You Score: Thinking Reward Model for Visual Generation

The Thinking Reward Model first sets case-specific rubrics and then scores, giving finer rewards for visual generation.

Affiliation unverified · Xuehai Bai, Zhenchen Tang, Yang Shi +9

Visual generation Alignment arXiv Code HF

▲103 ❝0 2026-09-29

Predictive Credit: Measuring What Scientific Explanations Add to Experimental Forecasts

Proposes 'predictive credit' to measure whether research agents' explanations improve experimental forecasts; preregistered tests were inconclusive.

Carnegie Mellon University · Jingjie Ning, Xueqi Li, Yibo Kong +1

AI for Science Benchmarks & data arXiv HF

▲101 ❝0 2026-09-29

Follow the Entities: A Corpus Map for Agentic Search

Microsoft builds an entity-based corpus map so search agents can follow evidence across documents instead of rediscovering links.

Microsoft · Soyeong Jeong, Sujay Kumar Jauhar, Sung Ju Hwang +1

Agents LLMs arXiv HF

▲93 ★29 ❝0 2026-09-29

SAKI: Maximal-Coupling-Routed Teacher Supervision for On-Policy Distillation

SAKI routes distillation supervision using KL-constrained teacher-guided rollouts and maximal coupling to help weak students.

meituan · Miteto Wei, Xiaohan Wang, Zehao Chen +7

Efficient training & inference arXiv Code HF

▲86 ★11 ❝1 2026-09-29

Learning Meta-Skills for Agent Harness Design in Test-Time AI4AI

Has a Builder model learn meta-skills to construct better harnesses for a fixed Target model at test time (AI-for-AI).

Apodex · Cheng Qian, Kunlun Zhu, Beibin Li +2

Agents arXiv Code HF

▲78 ★38 ❝2 2026-09-29

Imagine3D-LLM: Teaching MLLMs to Imagine 3D Scenes Before Answering

Imagine3D-LLM teaches MLLMs to imagine a 3D scene before answering multi-view spatial questions.

KAIST AI · Jaewoo Jung, Hyeonseo Yu, Honggyu An +10

Multimodal Reasoning arXiv Code HF

▲75 ★27 ❝0 2026-09-29

EVOKE: Eliciting World Knowledge in Agents for Transferable Decision-Making

EVOKE elicits world knowledge already learned in pretraining so LLM agents transfer better to unseen environments.

Shanghai Jiao Tong University · Yuhan Guo, Jinming Liu, Liang Xu +8

Agents arXiv Code HF

▲75 ★46 ❝0 2026-09-29

EngiWorld: What Can Frontier Agents Deliver in Professional Engineering Environments?

EngiWorld has 1,301 expert tasks across six engineering domains (CAD, CAE, EDA…) and 26 professional software platforms.

Zhiman AI · Hongcheng Gao, Hailong Qu, Yu Lei +19

Agents Benchmarks & data arXiv Code HF

▲68 ★10 ❝0 2026-09-29

E-MoE: Enhanced Mixture-of-Experts for Non-Factorized Diffusion Language Models

E-MoE improves non-factorised diffusion language models with enhanced mixture-of-experts, raising few-step quality without posterior collapse.

Affiliation unverified · Arseny Ivanov, Alexander Kolesov, Alexander Korotin +2

LLMs Efficient training & inference arXiv Code HF

▲64 ★18 ❝0 2026-09-29

In-context Robot Learning Made Simple: A Democratized Recipe for Manipulation Tasks

Clearly defines what robot in-context learning should follow and offers a minimalist, reproducible recipe for manipulation.

Affiliation unverified · Minxing Li, Minghao Han, Weizhi Zhao +10

Robotics & embodied AI arXiv Code HF

▲57 ★17 ❝0 2026-09-29

OmniTaskonomy: When Does Visual Generation Improve Visual Understanding?

Studies when visual generation training improves understanding, finding image-to-image training helps under the right recipe.

UC Berkeley · Jiaxin Ge, Yiming Qin, Ji Xie +13

Multimodal Vision & 3D arXiv Code HF

▲53 ❝2 2026-09-29

AIM: Agentic Idea Management for Automated Research

Google's AIM manages, selects and explores research ideas for automated research, inspired by Bayesian optimisation.

Google · Hyeong Kyu Choi, Bhavana Dalvi Mishra, Jiefeng Chen +7

Agents AI for Science arXiv HF

▲49 ★54 ❝1 2026-09-29

EVO-WAM: Evolving World Action Models through Video-Action Verification

EVO-WAM adapts world-action models to new tasks by learning from their own verified generated video-action trajectories.

Harbin Institute of Techonology, Shenzhen · Shiyang Zhou, Xionghao Wu, Wenbo Li +11

Robotics & embodied AI arXiv Code HF

▲47 ★742 ❝2 2026-09-29

Context Language Models

Context Language Models treat their context as an editable file and decide what to keep, gaining accuracy with fewer FLOPs.

Meta · Rulin Shao, Shannon Zejiang Shen, Junjie Oscar Yin +10

LLMs Agents arXiv Code HF

▲40 ★153 ❝0 2026-09-29

TERRA: Terrain-Aware Reconstruction, Retargeting and Control for Musculoskeletal Locomotion

TERRA retargets motion with terrain to musculoskeletal models so muscle-driven agents can walk on complex ground.

Mathis Group @ EPFL · Merkourios Simos, Chengkun Li, Bianca Ziliotto +1

Robotics & embodied AI Reinforcement learning arXiv Code HF

▲38 ★625 ❝0 2026-09-29

HelixWorld: A Real-time Interactive Audio-Visual World Model

HelixWorld is a real-time interactive world model where visuals and spatial stereo sound co-evolve with user input.

NoizAI · Lei Ke, Jiahao Pan, Zeyue Tian +13

Visual generation Multimodal arXiv Code HF

▲34 ★227 ❝1 2026-09-29

PixelUMM: Encoder-Free Unified Image and Video Understanding and Generation

PixelUMM unifies image and video understanding and generation directly in pixel space without a visual encoder.

NVIDIA · Cong Wei, Xuanchi Ren, Bryan Chu +8

Multimodal Visual generation arXiv Code HF

▲34 ★637 ❝9 2026-09-29

Systematically Exploring the Capabilities of GPT-6 Astra as Embodied Policies

Systematically tests GPT-6 Astra as an embodied policy across six domains, alone and combined with learned policies.

Galbot · Galbot Team, Xuchuan Chen, Xiaoqian Cheng +31

Robotics & embodied AI Benchmarks & data arXiv Code HF

▲21 ★2.8k ❝22 2026-09-29

TabFM: A Zero-Shot Foundation Model for Tabular Data

TabFM is a 400M-parameter tabular foundation model trained on synthetic tables that predicts zero-shot in a single forward pass.

Google · Weihao Kong, Erez Louidor Ilan, Shuxin Nie +6

LLMs Benchmarks & data arXiv Code HF

▲388 ★116 ❝2 2026-09-28

In-Context Learning for Robots: Methods and Applications

A survey of in-context learning for robots, grouping methods into four families by how context is turned into action.

Knowin AI · Haojian Huang, Zexi Li, Junhao Guo +36

Robotics & embodied AI arXiv Code HF

▲367 ★36 ❝1 2026-09-28

MaLiang-Harness: A Programmable Path to Image and Video Generation

MaLiang-Harness has MLLMs generate images and videos via programs in a build-inspect-revise loop that closes the program-to-visual gap.

National University of Singapore · Haoyu Zhao, Zihao Zhang, Xudong Wang +4

Visual generation Code & SWE arXiv Code HF

▲224 ★47 ❝0 2026-09-28

Rethinking Latent Visual Reasoning: Grounding Latent Reasoning in Visual Evidence

Shows latent visual reasoning tokens barely respond to image evidence and proposes grounding them in visual evidence during training.

Amazon · Xi Xiao, Tianchen Zhao, Youngeun Kim +10

Reasoning Multimodal arXiv Code HF

▲205 ❝1 2026-09-28

On-Policy or Off-Policy Learning? A Systematic Study of Distillation Dynamics

A controlled study isolating on- vs off-policy rollouts in distillation, finding a more nuanced picture than 'on-policy is always better'.

University of Cambridge · Julianna Piskorz, Antonin Berthon, Mihaela van der Schaar

Efficient training & inference Reasoning arXiv HF

▲181 ★57 ❝1 2026-09-28

Beyond Teacher Assignment: Domain-Normalized Multi-Teacher On-Policy Distillation

Normalises each teacher's feedback strength by domain in multi-teacher on-policy distillation so one student learns maths, code and instruction following.

Nanyang Technological University · Xin Li, Hao Jiang, Xin Gao +6

Efficient training & inference Reinforcement learning arXiv Code HF

▲151 ★70 ❝0 2026-09-28

PanoVLN: Towards Effective Panoramic Vision-and-Language Navigation

PanoVLN studies how to make vision-and-language navigation actually benefit from panoramic observations.

Zhejiang University · Zhen Wang, Changpeng Wang, Zhe Liu +5

Robotics & embodied AI Multimodal arXiv Code HF