ARCHIVE · 2026-10 · 198 PAPERS · PAGE 3/4

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)

▲58 ★38 ❝0 2026-10-12

AReA: Adaptive Receptive-Field Attention for Efficient High-Resolution Diffusion Transformers

AReA adapts the attention receptive field across denoising steps so diffusion transformers generate beyond-training resolutions more cheaply.

University of Toronto Computer Science · Javad Rajabi, Koorosh Roohi, David B. Lindell +1

Visual generation Efficient training & inference arXiv Code HF

▲24 ★1.1k ❝0 2026-10-12

StartLux-Decision: AutoThink Decision Models with Reinforcement Learning for Calibrated Deliberative Decisions

StartLux-Decision models (0.8B–27B) answer confident cases directly and reason only when uncertain, with RL that rewards calibrated decision probabilities.

StartLux · Xialie Zhuang, Zunhai Su, Jia Pan +3

Reasoning Reinforcement learning arXiv Code HF

▲143 ★30 ❝0 2026-10-10

Learning Where to Split: Dynamic Band Splitting for Audio Source Separation

Dynamic Band Splitting adapts frequency band boundaries per mixture to improve audio source separation within a fixed token budget.

Tsinghua Machine Learning Group · Bingzhe Zhang, Kai Li, Xiaolin Hu

Multimodal arXiv Code HF

▲88 ★30 ❝0 2026-10-10

Vela: User Behavior as Supervision Sparks LLM Intelligence

Vela uses real long-term user behaviour records as supervision to improve LLMs at inferring preferences, intent and future actions.

AllSpark Research · Yang Xiao, Yusong Sun, Haoyi Wu +17

LLMs arXiv Code HF

▲76 ★56 ❝0 2026-10-10

Last Secure Code Benchmark: Is the Code You Vibe Secure?

Builds 450 tasks from 150 real vulnerabilities to check whether code written by vibe-coding agents is secure.

Qwen · Tianyi Wu, Chao Ge, Peng Chen +12

Code & SWE Safety & interpretability Benchmarks & data arXiv Code HF

▲46 ★22 ❝0 2026-10-10

BridgeWAM: Bridging Video and Action Experts via Latent Bridge Queries

BridgeWAM links video and action experts through 32 latent bridge queries instead of dense layer-wise coupling, enabling a two-layer action head.

Fudan University · Jiaxin Fang, Ruitong Liu, Zeyu Lai +6

Robotics & embodied AI Multimodal arXiv Code HF

▲270 ★175 ❝0 2026-10-09

S-Space: Exploring Spatial Workspace in Multimodal Language Models

Finds S-Space, a linear subspace inside MLLMs where object tokens encode continuous coordinates, acting as an internal spatial workspace.

MirroS · Zhiyuan Shi, Xiang Li, Xingzhuo Guo +8

Multimodal Safety & interpretability arXiv Code HF

▲251 ❝0 2026-10-09

Beyond Domain-Routed Supervision: Cross-Teacher Guarding for Multi-Teacher On-Policy Distillation

Adds cross-teacher guarding to multi-teacher on-policy distillation to reduce gradient conflicts between domain teachers.

jingdong · Shi-Yu Tian, Yong Liu, Ximan Liu +10

Efficient training & inference arXiv HF

▲233 ★42 ❝0 2026-10-09

AgentSWE: Can Coding Agents Build the Agent You Actually Want?

AgentSWE benchmarks whether coding agents can create, edit and optimise agents that truly meet the stated requirement.

Affiliation unverified · Jiahao Wang, Hongjin Qian, Yuyang Hu +2

Agents Code & SWE Benchmarks & data arXiv Code HF

▲165 ❝0 2026-10-09

Scaling Agentic Data for Long-Horizon Terminal Intelligence

Scales synthesis of high-quality long-horizon terminal trajectories to train agents for professional tasks.

alibaba · Shun Zou, Ziyu Ma, Yi Zou +6

Agents Benchmarks & data arXiv HF

▲115 ★14k ❝1 2026-10-09

Harness^2: Recursive Agent Harnessing for an Open World

Harness² lets agents recursively improve their own harness on unseen open-world tasks without ground-truth verifiers.

Google · Ruiyao Xu, Yanfei Chen, Zhongying CuiZhu +6

Agents arXiv Code HF

▲74 ★17 ❝0 2026-10-09

EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents

EmbodiedCLBench is the first embodied continual-learning benchmark, testing whether agents recombine learned skills on new tasks.

PKU-VaLuE-Lab · Jie Huang, Yanan Chen, Ruixun Liu +6

Robotics & embodied AI Agents Benchmarks & data arXiv Code HF

▲55 ★33 ❝1 2026-10-09

OpenMobile-2: Building Versatile Mobile Agents with Scalable Environments and App-Native Tools

OpenMobile-2 releases environments spanning 110+ apps and ~12K trajectories to train a near-frontier, fully open mobile agent.

Nanjing University · Chenyang Yan, Yingying Zhang, Kanzhi Cheng +21

Agents Reinforcement learning arXiv Code HF

▲54 ★809 ❝2 2026-10-09

RSI-Anything: Unlocking Recursive Self-Improvement with World Models

RSI-Anything uses predictive world models to guide actions and data selection for domain-general recursive self-improvement.

PhAI Labs · Taoyong Cui, Pinxian Zeng, Wanghan Xu +8

Agents AI for Science arXiv Code HF

▲52 ★34 ❝0 2026-10-09

Work2Skill: Capturing Human GUI Experience for Reusable Agent Skill Induction

Work2Skill captures human GUI demonstrations and induces reusable skills for computer-use agents.

Nex AGI · Yuqi Huang, Hui Li, Yifei Cao +9

Agents Multimodal arXiv Code HF

▲50 ★37 ❝0 2026-10-09

Multi-Stream Attention Residuals

Multi-Stream Attention Residuals extend depth-wise attention residuals to multiple heads, with sizeable gains at little extra cost.

Shanghai Jiao Tong University · Xiaohan Qin, Xiangdong Zhang, Tuo Dai +7

LLMs Efficient training & inference arXiv Code HF

▲37 ★67 ❝0 2026-10-09

Correction Training for Diffusion Models

Trains a corrector that refines the denoiser's prediction at each step to reduce exposure bias, improving DiT, SiT and JiT samples.

University of California, Berkeley · Donghoon Ahn, Runqian Wang, Alexei A. Efros +2

Visual generation arXiv Code HF

▲33 ★16 ❝0 2026-10-09

Learning Representations to Guide Pixel Diffusion Transformers

Studies which visual representations actually help pixel diffusion transformers, finding better understanding does not imply better generation.

alibaba · Zihao Pan, Mengting Chen, Cheng Shi +6

Visual generation arXiv Code HF

▲26 ★4 ❝0 2026-10-09

APEX-Harm: Evaluating Agent Safety in Shared Professional Workspaces

APEX-Harm tests finance, consulting and legal agents against prompt and script injection in large shared workspaces.

King Abdullah University of Science and Technology · Ziyi Yang, Adel Bibi, Guohao Li +2

Safety & interpretability Agents Benchmarks & data arXiv Code HF

▲325 ★154 ❝0 2026-10-08

DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models

DiffusionART aggregates rewards before tilting the objective, making online RL fine-tuning of diffusion models efficient and stable.

KlingAI Research · Zhenglin Wan, Jiayu Chen, Zhengfeiyang Zhang +6

Visual generation Reinforcement learning arXiv Code HF

▲293 ★201 ❝1 2026-10-08

AgentGarten: Code Worlds for Evolving Agents

AgentGarten couples simulators and game engines with a neural renderer to build faithful, realistic worlds for training agents.

MirroS · Jiawei Chi, Shangchen Miao, Zhiyuan Shi +12

Agents Visual generation arXiv Code HF

▲241 ★9 ❝0 2026-10-08

Thinking Inertia: LLMs Keep Thinking When Told Not To

Shows LLMs keep reasoning even when told not to think, and proposes a three-level way to measure no-thinking behaviour.

Stanford University · Dianqiao Lei, Kevin Qinghong Lin, Pan Lu +2

Reasoning LLMs arXiv Code HF

▲167 ★303 ❝0 2026-10-08

REMORY: Learning Residual Memory for Context Compaction

REMORY appends learned soft memory tokens after a summary so long-horizon agents lose less when compacting context.

Fusion Lab: Generative Vision Lab of Fudan University and SII · Hanchen Xia, Baoyou Chen, Yutang Ge +5

Agents Efficient training & inference arXiv Code HF

▲154 ★34 ❝0 2026-10-08

Reflex: Visually Grounded Reactive Humanoid Control

Reflex enables humanoids to react with whole-body control from egocentric vision, such as catching a thrown box.

University of Washington · Taoyang Jia, Weikai Huang, Linxin Song +6

Robotics & embodied AI Reinforcement learning arXiv Code HF

▲143 ★63 ❝0 2026-10-08

TokenRouter: Efficient Serving System for Token-Level LLM Routing

TokenRouter is a serving system for token-level LLM routing that fixes step desynchronisation and batching delays.

Tsinghua-NICS-EFC · Tianyu Fu, Tengxuan Liu, Ruoxi Wang +4

Efficient training & inference arXiv Code HF

▲111 ★20 ❝0 2026-10-08

SoftwareGym: Scaling Computer-Use Agent Training with Diverse Synthesized Software Environments

SoftwareGym automatically synthesises diverse, executable, verifiable GUI software environments to scale computer-use agent training.

Shanghai Jiao Tong University · Tian Jin, Wenhao Wang, Jing Kang +4

Agents Benchmarks & data arXiv Code HF

▲92 ❝6 2026-10-08

MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement

Technical report of Xiaomi's omni-modal MiMo-V2.6, focused on scaling RL compute along several dimensions toward self-improvement.

Xiaomi MiMo · Xiaomi LLM-Core Team, :, Zongming Qiao +147

LLMs Reinforcement learning Multimodal arXiv HF

▲86 ★198 ❝0 2026-10-08

SuperNav: An Agentic Navigation System for Any Task in Any Scene

SuperNav lets an MLLM interpret requests and decide while delegating motion to navigation tools, generalising across tasks and scenes.

zju3dv · Jinkai Zhang, Jingyi Xu, Yuanhong Yu +5

Robotics & embodied AI Agents arXiv Code HF

▲75 ★287 ❝0 2026-10-08

LeWAM: End-to-End World and Action Modeling with JEPAs

LeWAM is an end-to-end JEPA world-action model that supports both direct policy execution and efficient planning.

LeWAM · Minghao Fu, Tavis Siebert, Eryk Halicki +1

Robotics & embodied AI Vision & 3D arXiv Code HF

▲72 ★21 ❝1 2026-10-08

Multi-Agent Egocentric World Model with Fine-Grained Embodied Interaction

A multi-agent egocentric world model that generates synchronised first-person streams of agents interacting in a shared world.

KAIST AI · Dahyun Chung, Siyoon Jin, Hyunwook Choi +5

Visual generation Robotics & embodied AI arXiv Code HF

▲64 ★44 ❝1 2026-10-08

OuroWorld: Bringing Any 3D World Alive as Diverse, Endlessly Looping 3D Cinemagraphs

OuroWorld turns any static 3D Gaussian scene into a seamlessly looping 3D cinemagraph viewable from any angle.

Alaya Lab · You-Zhe Xie, Ting-Wei Chou, Yu-Hsuan Li +3

Visual generation arXiv Code HF

▲197 ★123 ❝0 2026-10-07

UltraText Bench: A Comprehensive Bilingual Benchmark for Evaluating Visual Text Rendering in Image Generation

UltraText Bench uses 432 bilingual prompts to evaluate dense visual text rendering in image generators.

Westlake University · Deyuan Liu, Yihao Hu, Jingxuan Zhang +13

Visual generation Benchmarks & data arXiv Code HF

▲132 ★9 ❝0 2026-10-07

LiveMACE: Process-Aware Evaluation of LLM Agent Capabilities in Evolving Markets

LiveMACE uses live financial markets to evaluate persistent LLM agents on both outcomes and process diagnostics.

Fudan University · Jun Zhao, Leiming Fu, Yanbo Wen +9

Agents Benchmarks & data arXiv Code HF

▲113 ★2.8k ❝1 2026-10-07

Long-WAM: Scaling the Context of World-Action Models

Long-WAM scales the context of world-action models under real-time limits and finds autoregressive video pretraining makes long history pay off.

NVIDIA · Wei Huang, Bohan Zhang, Chenzhi Liu +13

Robotics & embodied AI Efficient training & inference arXiv Code HF

▲102 ★12 ❝0 2026-10-07

VIS-Ground: Video Interactive Storytelling with Contextual Grounding

VIS-Ground lets viewers steer interactive video stories while staying consistent with the source and rendered history.

Google · Bingxuan Li, Yiwen Song, Xueqing Wu +10

Visual generation arXiv Code HF

▲90 ★26 ❝0 2026-10-07

GRACE: Generation-aware latent compression for efficient video generation

GRACE further compresses the video latent while staying compatible with the pretrained DiT, speeding up video generation.

KAIST AI · Jiyoung Kim, Paul Hyunbin Cho, Jisu Nam +5

Visual generation Efficient training & inference arXiv Code HF

▲60 ★207 ❝0 2026-10-07

SGF+: Decoupling Gradient Flows for Autoregressive Video Generation

SGF+ separates parameters for context writing and denoising in autoregressive video generation to resolve conflicting gradients.

Affiliation unverified · Zihan Su, Junhao Zhuang, Yaowei Li +10

Visual generation arXiv Code HF

▲49 ★41 ❝0 2026-10-07

Tetris3D: 3D Scene Generation With Objects That Fit Together

Tetris3D conditions each object on its neighbours' geometry and physical relations so single-image 3D scenes fit together.

KAIST AI · Jaeyeong Kim, Jinhyuk Jang, Jongmin Lee +2

Vision & 3D arXiv Code HF

▲45 ★19 ❝1 2026-10-07

RLHND: Video Foundation Models as Physically Grounded Hand Trackers for Robot Learning

RLHND uses a video foundation model to estimate hand pose and touch from egocentric video, making human videos more useful for robots.

RLWRLD · Seungjun Moon, Subin Jeon, Sangwoo Kim +2

Robotics & embodied AI Vision & 3D arXiv Code HF

▲40 ★41 ❝1 2026-10-07

RobotWorld: Benchmarking Multimodal Agents for Robot Use Across Diverse Tasks and Embodiments

RobotWorld uses 84 simulated tasks (manipulation, locomotion, driving, aerial) to test multimodal agents at robot use.

Affiliation unverified · Zhiqin Yang, Chenxin Li, Xiaomeng Hu +30

Robotics & embodied AI Agents Benchmarks & data arXiv Code HF

▲31 ★194 ❝0 2026-10-07

Iris-3B: Going Beyond the Latent with Pixel-Space Diffusion Training, Conversion and Fine-Tuning

Iris-3B pretrains and converts pixel-space diffusion models to test whether skipping the VAE helps detail-sensitive tasks.

speridlabs · Hanqiu Li Cai, Chema Garabito

Visual generation arXiv Code HF

▲189 ★10 ❝1 2026-10-06

Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight

Self-Retrospection Distillation turns hindsight into foresight supervision, filling the gap when RLVR rewards are all identical.

Salesforce AI Research · Haoxiang Zhang, Qinglin Chen, Hiroaki Hayashi +9

Reinforcement learning Agents arXiv Code HF

▲153 ★13 ❝1 2026-10-06

Learn2Play Bench: How Well Do LLM Agents Learn from Experience in Unfamiliar Environments?

Learn2Play Bench uses text games with novel or counterintuitive rules to test whether agents learn from experience.

National University of Singapore · Yibo Li, Jinhang Qiu, Zhi Zheng +4

Agents Benchmarks & data arXiv Code HF

▲116 ★1.6k ❝0 2026-10-06

nanoMuse: An Open-Source Personal Agent for Every Device You Own

nanoMuse is an open-source personal agent that spans a user's devices and accounts, positioned as an open counterpart to Meta's Muse.

Zhejiang University · Guangyi Liu, Yong Liu, Jiangning Zhang

Agents arXiv Code HF

▲97 ★74 ❝1 2026-10-06

Recursive Game Creator: An Agentic Product-Level Experience-Oriented Game Harness

Recursive Game Creator iterates with designer, builder, player and reviewer roles so agents make games that are actually fun.

The University of Hong Kong · Jiajun Chen, Haoyu Wu, Mingda Jia +1

Agents Code & SWE arXiv Code HF

▲91 ❝0 2026-10-06

TRACE: Rollout-Guided Quantization-Aware Training for FP4 Reinforcement Learning of MoE Language Models

Qwen's TRACE aligns the training and rollout FP4 quantisation paths for low-precision RL of MoE language models.

Qwen · Xin Wang, Hao Yu, Zhengyang Zhuge +9

Efficient training & inference Reinforcement learning LLMs arXiv HF

▲85 ★5 ❝0 2026-10-06

DecepEval: A Benchmark for Evaluating Deception in LLM Agents

DecepEval uses 1,532 instances and a 'deception diamond' framework to test when LLM agents are likely to deceive.

Xi'an Jiaotong University · Yiming Xu, Hongyue Yu, Beihua Yang +8

Safety & interpretability Benchmarks & data Agents arXiv Code HF

▲72 ★2 ❝2 2026-10-06

CheckerBench: Can Long-Horizon Agents Synthesize Static-Analysis Checkers?

CheckerBench tests whether long-horizon agents can build working static-analysis checkers, with 300 tasks from 297 CVEs.

Humanlaya Data Lab · Hang He, Li Wang, Hao Chen +9

Code & SWE Agents Benchmarks & data arXiv Code HF

▲70 ★91 ❝0 2026-10-06

VisionWeave: Weaving Elastic Visual Representations as a Native Capability of MLLMs

VisionWeave lets MLLMs learn end-to-end where and how finely to spend visual tokens, adapting to image content.

Qwen · Yuan Feng, Qize Yang, Ruizhe Chen +9

Multimodal Efficient training & inference arXiv Code HF

▲54 ★11 ❝0 2026-10-06

U-Space: Uncovering When and Why Uncertainty Arises in Language Models

U-Space analyses when and why uncertainty arises in language models, to help judge whether an individual answer can be trusted.

Multimodal AI Lab - Technical University of Darmstadt · Tobias Braun, Nils Loose, Alexander Herzog +4

Safety & interpretability LLMs arXiv Code HF