Monthly collection

May 2026

The papers that drew the most attention during this month of AI research.

Papers: 115
Citations: 503
Research labs: 5

Explore the year

View all months →

Aug2025 Sep2025 Oct2025 Nov2025 Dec2025 Jan2026 Feb2026 Mar2026 Apr2026 May2026 Jun2026 Jul2026 Aug2026

1,016papers
curated and indexed

115 papers published in May 2026

01

NVIDIA

Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players

The paper introduces Gamma-World, a generative multi-agent world model for interactive simulation that scales beyond two players. It addresses limitations of prior work like Solaris, which uses dense attention and learned per-slot identities, by proposing two key innovations. First, Simplex Rotary Agent Encoding extends 3D RoPE with an agent axis,…

Fangfu Liu, Kai He, Tianchang Shen, Tianshi Cao, et al.

Published: May 2026
Upvotes: 433
Citations: 2

Read summary Original

02

arXiv.org

MolmoAct2: Action Reasoning Models for Real-world Deployment

MolmoAct2 is a fully open Vision-Language-Action (VLA) model family for real-world robot deployment, built on the Molmo2-ER backbone, a VLM specialized for spatial and embodied reasoning trained on a 3.3M-sample corpus. It introduces three new robot datasets: MolmoAct2-BimanualYAM (720 hours, largest open bimanual dataset), MolmoAct2-DROID…

Haoquan Fang, Jiafei Duan, Donovan Clay, Sam Wang, et al.

Published: May 2026
Upvotes: 356
Citations: 23

Read summary Original

03

arXiv.org

HRM-Text: Efficient Pretraining Beyond Scaling

HRM-Text introduces a Hierarchical Recurrent Model (HRM) that decouples computation into slow strategic and fast execution layers, inspired by biological multi-timescale processing. Trained from scratch on only 40B unique tokens (60B total with repetition) and a $1,500 budget, a 1B-parameter model achieves 60.7% on MMLU, 81.9% on ARC-C, 82.2% on DROP,…

Guan Wang, Changling Liu, Chenyu Wang, Cai Zhou, et al.

Published: May 2026
Upvotes: 323
Citations: 3

Read summary Original

04

arXiv.org

CiteVQA: Benchmarking Evidence Attribution for Trustworthy Document Intelligence

CiteVQA is a benchmark for evaluating Multimodal Large Language Models (MLLMs) on document visual question answering, requiring both correct answers and element-level bounding-box citations. It includes 1,897 questions from 711 PDFs across seven domains and two languages, with an average of 40.6 pages per document. Ground-truth citations are generated via…

Dongsheng Ma, Jiayu Li, Zhengren Wang, Yijie Wang, et al.

Published: May 2026
Upvotes: 274
Citations: 3

Read summary Original

05

arXiv.org

SkillOpt: Executive Strategy for Self-Evolving Agent Skills

SkillOpt is a text-space optimizer that treats an agent's skill document as trainable external state, using a separate optimizer model to propose bounded add/delete/replace edits based on scored rollouts. Edits are accepted only if they improve a held-out validation score, with controls like a textual learning-rate budget, rejected-edit buffer, and…

Yifan Yang, Ziyang Gong, Weiquan Huang, Qihao Yang, et al.

Published: May 2026
Upvotes: 264
Citations: 29

Read summary Original

06

arXiv.org

Crafter: A Multi-Agent Harness for Editable Scientific Figure Generation from Diverse Inputs

The paper introduces CRAFTER, a multi-agent harness for generating scientific figures from diverse inputs, and CRAFTEDITOR, which converts raster outputs into editable SVGs. CRAFTER uses five agents (intent reasoner, plan generator, critic, specification refiner, convergence judge) sharing an evolving specification, with mechanisms for diversity-driven…

Haozhe Zhao, Shuzheng Si, Zhenhailong Wang, Zheng Wang, et al.

Published: May 2026
Upvotes: 253
Citations: 1

Read summary Original

07

arXiv.org

Mean Mode Screaming: Mean--Variance Split Residuals for 1000-Layer Diffusion Transformers

This paper identifies a failure mode in ultra-deep Diffusion Transformers (DiTs) called Mean Mode Screaming (MMS), where token representations homogenize and centered variation is suppressed, leading to training collapse. The mechanism involves a mean-coherent gradient shock that opens residual branches, causing a mean-dominated state. The authors propose…

Pengqi Lu

Published: May 2026
Upvotes: 238
Citations: 0

Read summary Original

08

arXiv.org

MinT: Managed Infrastructure for Training and Serving Millions of LLMs

MinT is a managed infrastructure system for Low-Rank Adaptation (LoRA) post-training and online serving, designed for settings where many trained policies are produced over a small number of expensive base-model deployments. Instead of materializing full checkpoints, MinT keeps the base model resident and moves exported LoRA adapter revisions through…

Mind Lab, :, Song Cao, Vic Cao, et al.

Published: May 2026
Upvotes: 225
Citations: 1

Read summary Original

09

arXiv.org

Code as Agent Harness

This survey introduces the concept of 'code as agent harness,' reframing code from a mere output of large language models (LLMs) to the operational substrate for agentic AI systems. It argues that code serves as an executable, inspectable, and stateful medium through which agents reason, act, and model their environments. The paper organizes the literature…

Xuying Ning, Katherine Tieu, Dongqi Fu, Tianxin Wei, et al.

Published: May 2026
Upvotes: 225
Citations: 12

Read summary Original

10

arXiv.org

DelTA: Discriminative Token Credit Assignment for Reinforcement Learning from Verifiable Rewards

The paper introduces DelTA, a method for improving reinforcement learning from verifiable rewards (RLVR) in large language models. The authors show that sequence-level RLVR updates act as an implicit linear discriminator over token-gradient vectors, determining which token probabilities increase or decrease. Standard RLVR forms this discriminator using…

Kaiyi Zhang, Wei Wu, Yankai Lin

Published: May 2026
Upvotes: 207
Citations: 0

Read summary Original

11

arXiv.org

Anti-Self-Distillation for Reasoning RL via Pointwise Mutual Information

The paper identifies a structural bias in on-policy self-distillation for math reasoning, where the per-token signal (conditional pointwise mutual information between the next token and privileged context) rewards shortcut tokens (e.g., 'Given', 'succeeds') and penalizes deliberation tokens (e.g., 'Wait', 'Let'). To fix this, the authors propose…

Guobin Shen, Xiang Cheng, Chenxiao Zhao, Lei Huang, et al.

Published: May 2026
Upvotes: 196
Citations: 6

Read summary Original

12

arXiv.org

SenseNova-U1: Unifying Multimodal Understanding and Generation with NEO-unify Architecture

SenseNova-U1 is a native unified multimodal model built on the NEO-unify architecture, designed to overcome the traditional divide between understanding and generation. It operates directly on raw pixels and text, eliminating the need for pretrained vision encoders (VEs) and variational autoencoders (VAEs). The model uses a near-lossless visual interface…

Haiwen Diao, Penghao Wu, Hanming Deng, Jiahao Wang, et al.

Published: May 2026
Upvotes: 195
Citations: 12

Read summary Original

13

arXiv.org

AutoResearchClaw: Self-Reinforcing Autonomous Research with Human-AI Collaboration

AutoResearchClaw is a multi-agent autonomous research pipeline that addresses three key challenges in scientific discovery: hypothesis quality, execution robustness, and experience accumulation. It integrates five mechanisms: structured multi-agent debate for hypothesis generation and result analysis, a self-healing executor with a Pivot/Refine loop,…

Jiaqi Liu, Shi Qiu, Mairui Li, Bingzhou Li, et al.

Published: May 2026
Upvotes: 191
Citations: 13

Read summary Original

14

arXiv.org

TransitLM: A Large-Scale Dataset and Benchmark for Map-Free Transit Route Generation

TransitLM is a large-scale dataset of over 13 million transit route planning records from four Chinese cities (Beijing, Shanghai, Shenzhen, Chengdu), covering 120,845 stations and 13,666 lines. It is released as a continual pre-training corpus and benchmark data for three tasks: optimal route generation, preference-aware planning, and multi-route…

Hanyu Guo, Jiedong Yang, Chao Chen, Longfei Xu, et al.

Published: May 2026
Upvotes: 179
Citations: 0

Read summary Original

15

arXiv.org

Perception or Prejudice: Can MLLMs Go Beyond First Impressions of Personality?

This paper introduces Grounded Personality Reasoning (GPR) and the MM-OCEAN benchmark to evaluate whether Multimodal Large Language Models (MLLMs) perceive personality through behavioral evidence or merely prejudge via superficial patterns. The benchmark includes 1,104 videos and 5,320 cue-grounding MCQs, built via a multi-agent human-collaborative…

Caixin Kang, Tianyu Yan, Sitong Gong, Mingfang Zhang, et al.

Published: May 2026
Upvotes: 171
Citations: 0

Read summary Original

16

arXiv.org

Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling

The paper introduces SU-01, a 30B-A3B reasoning model achieving gold-medal-level performance on mathematical and physical olympiads through a simple, unified post-training recipe. The recipe comprises three stages: (1) SFT with a reverse-perplexity curriculum on 338K rigorous proof trajectories to instill proof-search and self-checking behaviors; (2) a…

Yafu Li, Runzhe Zhan, Haoran Zhang, Shunkai Zhang, et al.

Published: May 2026
Upvotes: 166
Citations: 4

Read summary Original

17

arXiv.org

When Vision Speaks for Sound

The paper identifies a 'Clever Hans effect' in video-capable multimodal LLMs, where models appear to understand audio but actually rely on visual cues to hallucinate or infer sounds without verifying the audio stream. This is demonstrated across open-source and closed-source models. To systematically study this, the authors introduce THUD, a diagnostic…

Xiaofei Wen, Wenjie Jacky Mo, Xingyu Fu, Rui Cai, et al.

Published: May 2026
Upvotes: 161
Citations: 2

Read summary Original

18

arXiv.org

Domino: Decoupling Causal Modeling from Autoregressive Drafting in Speculative Decoding

Domino is a speculative decoding framework that decouples causal dependency modeling from expensive autoregressive draft execution. It uses a parallel draft backbone (based on DFlash) to generate preliminary draft distributions for the entire block, then applies a lightweight Domino head—comprising a GRU causal encoder and a low-rank correction head—to…

Jianuo Huang, Yaojie Zhang, Qituan Zhang, Hao Lin, et al.

Published: May 2026
Upvotes: 152
Citations: 8

Read summary Original

19

arXiv.org

MemPrivacy: Privacy-Preserving Personalized Memory Management for Edge-Cloud Agents

MemPrivacy is a framework for privacy-preserving memory management in edge-cloud LLM agents. It detects privacy-sensitive spans on edge devices, replaces them with type-aware placeholders (e.g., <Health_Info_1>) before cloud processing, and restores original values locally. This decouples privacy protection from semantic destruction, preserving…

Yining Chen, Jihao Zhao, Bo Tang, Haofen Wang, et al.

Published: May 2026
Upvotes: 150
Citations: 7

Read summary Original

20

Research paper

AgentDoG 1.5: A Lightweight and Scalable Alignment Framework for AI Agent Safety and Security

AgentDoG 1.5, developed by Shanghai Artificial Intelligence Laboratory, is a lightweight and scalable framework for AI agent safety and security. It updates the three-dimensional safety taxonomy (risk source, failure mode, real-world harm) to cover new risks from Codex and OpenClaw execution scenarios, extending the ATBench benchmark family with…

Dongrui Liu, Yu Li, Zhonghao Yang, Peng Wang, et al.

Published: May 2026
Upvotes: 150
Citations: 2

Read summary Original

21

arXiv.org

Video2GUI: Synthesizing Large-Scale Interaction Trajectories for Generalized GUI Agent Pretraining

The paper introduces Video2GUI, a fully automated framework that extracts grounded GUI interaction trajectories from unlabeled internet videos to address the scarcity of large-scale training data for GUI agents. The pipeline uses a coarse-to-fine filtering strategy: first, metadata-based classification (using a fine-tuned Qwen2.5-7B) reduces 500 million…

Weimin Xiong, Shuhao Gu, Bowen Ye, Zihao Yue, et al.

Published: May 2026
Upvotes: 147
Citations: 2

Read summary Original

22

NVIDIA

LocateAnything: Fast and High-Quality Vision-Language Grounding with Parallel Box Decoding

LocateAnything is a unified vision-language model for visual grounding and detection that introduces Parallel Box Decoding (PBD). Unlike standard next-token prediction (NTP) which serializes bounding box coordinates into 1D token streams, PBD treats each bounding box as an atomic unit, predicting all its coordinates in a single forward pass. This…

Shihao Wang, Shilong Liu, Yuanguo Kuang, Xinyu Wei, et al.

Published: May 2026
Upvotes: 146
Citations: 7

Read summary Original

23

Qwen

Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments

Qwen-VLA is a unified embodied foundation model that extends the Qwen3.5-4B vision-language backbone with a DiT-based flow-matching action decoder, enabling a single model to handle manipulation, navigation, and trajectory prediction across diverse tasks, environments, and robot embodiments. It uses embodiment-aware prompt conditioning and a unified…

Qiuyue Wang, Mingsheng Li, Jian Guan, Jinhui Ye, et al.

Published: May 2026
Upvotes: 146
Citations: 20

Read summary Original

24

arXiv.org

PhysBrain 1.0 Technical Report

PhysBrain 1.0 introduces a training strategy for embodied AI that prioritizes physical commonsense acquisition over pure action imitation. Its data engine converts large-scale human egocentric video (Ego4D, BuildAI, EgoDex, EPIC, SEA-Small) into structured scene meta-information—scene elements, spatial dynamics, action execution, and depth-aware…

Shijie Lian, Bin Yu, Xiaopeng Lin, Changti Wu, et al.

Published: May 2026
Upvotes: 145
Citations: 0

Read summary Original

25

arXiv.org

ARIS: Autonomous Research via Adversarial Multi-Agent Collaboration

ARIS is an open-source research harness for autonomous ML research, built on the assumption that single-agent long-horizon tasks are unreliable. It uses cross-model adversarial collaboration: an executor model (e.g., Claude) drives progress while a reviewer from a different model family (e.g., GPT-5.4) critiques artifacts and requests revisions. The system…

Ruofeng Yang, Yongcan Li, Shuai Li

Published: May 2026
Upvotes: 144
Citations: 9

Read summary Original

26

arXiv.org

MulTaBench: Benchmarking Multimodal Tabular Learning with Text and Image

MulTaBench is a benchmark of 40 datasets (20 image-tabular, 20 text-tabular) designed for Multimodal Tabular Learning (MMTL). It addresses the limitation of existing benchmarks that focus on modality co-occurrence, which masks the benefits of task-specific tuning. The authors propose two criteria for dataset inclusion: Joint Signal (each modality provides…

Alan Arazi, Eilam Shapira, Shoham Grunblat, Mor Ventura, et al.

Published: May 2026
Upvotes: 142
Citations: 2

Read summary Original

27

arXiv.org

DVAO: Dynamic Variance-adaptive Advantage Optimization for Multi-reward Reinforcement Learning

The paper proposes Dynamic Variance-adaptive Advantage Optimization (DVAO) to improve multi-reward reinforcement learning for LLMs, addressing flaws in standard scalarization methods. Reward Combination (RC) causes training instability due to large advantage magnitudes, while Advantage Combination (AC) uses static weights and ignores cross-objective…

Guochao Jiang, Jingyi Song, Guofeng Quan, Chuzhan Hao, et al.

Published: May 2026
Upvotes: 139
Citations: 1

Read summary Original

28

arXiv.org

Mega-ASR: Towards In-the-wild^2 Speech Recognition via Scaling up Real-world Acoustic Simulation

MEGA-ASR is a framework for automatic speech recognition (ASR) in real-world environments, addressing the 'acoustic robustness bottleneck' where models fail under severe, compositional distortions. The authors introduce VOICES-IN-THE-WILD-2M, a large-scale dataset with 7 atomic acoustic effects (noise, far-field, obstructed, echo&reverb, recording,…

Zhifei Xie, Kaiyu Pang, Haobin Zhang, Deheng Ye, et al.

Published: May 2026
Upvotes: 137
Citations: 1

Read summary Original

29

arXiv.org

$δ$-mem: Efficient Online Memory for Large Language Models

δ-mem is a lightweight memory mechanism that augments a frozen full-attention LLM with a compact online state of associative memory, updated via delta-rule learning. It compresses past information into a fixed-size state matrix (e.g., 8×8) and uses its readout to generate low-rank corrections to the backbone's attention computation, avoiding full…

Jingdi Lei, Di Zhang, Junxian Li, Weida Wang, et al.

Published: May 2026
Upvotes: 133
Citations: 3

Read summary Original

30

arXiv.org

SkillsVote: Lifecycle Governance of Agent Skills from Collection, Recommendation to Evolution

SkillsVote is a framework for governing Agent Skills across collection, recommendation, attribution, and evolution. It profiles a corpus of over one million open-source skills for environment requirements, quality, and verifiability. Before task execution, it performs agentic library search to recommend relevant skills and generate usage guidance. After…

Hongyi Liu, Haoyan Yang, Tao Jiang, Bo Tang, et al.

Published: May 2026
Upvotes: 131
Citations: 14

Read summary Original

31

arXiv.org

Stream-R1: Reliability-Perplexity Aware Reward Distillation for Streaming Video Generation

Stream-R1 is a framework for distilling autoregressive streaming video diffusion models, addressing limitations in existing distribution matching distillation (DMD) methods that treat all rollouts, frames, and pixels equally. It introduces two concepts: Inter-Reliability (varying reliability of supervision across rollouts) and Intra-Perplexity (varying…

Bin Wu, Mengqi Huang, Shaojin Wu, Weinan Jia, et al.

Published: May 2026
Upvotes: 129
Citations: 2

Read summary Original

32

arXiv.org

RLDX-1 Technical Report

RLDX-1 is a general-purpose Vision-Language-Action model (VLA) for dexterous manipulation, integrating versatile intelligence with functional capabilities: motion awareness, long-term memory, and physical sensing. Built on the Multi-Stream Action Transformer (MSAT) and a Qwen3-VL 8B backbone, it processes heterogeneous modalities via dedicated streams with…

Dongyoung Kim, Huiwon Jang, Myungkyu Koo, Suhyeok Jang, et al.

Published: May 2026
Upvotes: 128
Citations: 11

Read summary Original

33

arXiv.org

Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction

The paper introduces Direct Corpus Interaction (DCI), a retrieval paradigm where agents search raw corpora using terminal tools (grep, bash, file reads) instead of conventional retrievers. DCI bypasses embedding models and vector indexes, enabling fine-grained, compositional search. Evaluated on BrowseComp-Plus, multi-hop QA, and IR benchmarks,…

Zhuofeng Li, Haoxiang Zhang, Cong Wei, Pan Lu, et al.

Published: May 2026
Upvotes: 127
Citations: 21

Read summary Original

34

arXiv.org

COLLEAGUE.SKILL: Automated AI Skill Generation via Expert Knowledge Distillation

COLLEAGUE.SKILL is an open-source system that automatically distills heterogeneous traces of a person or role (e.g., work documents, chat logs, public interviews) into portable, inspectable, and correctable AI skill packages. The system generates a versioned package with two tracks: a capability track for practices and decision heuristics, and a bounded…

Tianyi Zhou, Dongrui Liu, Leitao Yuan, Jing Shao, et al.

Published: May 2026
Upvotes: 124
Citations: 3

Read summary Original

35

arXiv.org

MMSkills: Towards Multimodal Skills for General Visual Agents

MMSkills is a framework for representing, generating, and using reusable multimodal procedural knowledge for visual agents. Each MMSkill package combines a textual procedure, runtime state cards (encoding when-to-use conditions, visible cues, verification cues), and multi-view keyframes (full-frame, focus crop, before/after) to support state-conditioned…

Kangning Zhang, Shuai Shao, Qingyao Li, Jianghao Lin, et al.

Published: May 2026
Upvotes: 123
Citations: 3

Read summary Original

36

Qwen

Qwen-Image-2.0 Technical Report

Qwen-Image-2.0 is an image generation foundation model that unifies text-to-image (T2I) generation and instruction-based image editing in a single framework. It addresses challenges in ultra-long text rendering (up to 1K tokens), multilingual typography, high-resolution photorealism (native 2K), complex instruction following, and inference efficiency. The…

Bing Zhao, Chenfei Wu, Deqing Li, Hao Meng, et al.

Published: May 2026
Upvotes: 118
Citations: 8

Read summary Original

37

arXiv.org

Self-Distilled Agentic Reinforcement Learning

The paper introduces SDAR (Self-Distilled Agentic Reinforcement Learning), a method for post-training multi-turn LLM agents. It addresses issues with On-Policy Self-Distillation (OPSD), which provides dense token-level guidance but suffers from multi-turn instability and asymmetric trust in privileged guidance (e.g., retrieved skills). SDAR keeps GRPO as…

Zhengxi Lu, Zhiyuan Yao, Zhuowen Han, Zi-Han Wang, et al.

Published: May 2026
Upvotes: 118
Citations: 13

Read summary Original

38

arXiv.org

GrepSeek: Training Search Agents for Direct Corpus Interaction

GrepSeek is a Direct Corpus Interaction (DCI) search agent that bypasses traditional retrieval indexes by having a compact LLM issue Unix shell commands (e.g., rg, grep) directly over a raw text corpus. It uses a two-stage training pipeline: first, a cold-start dataset is generated via an answer-aware Tutor (backward chaining with answer-leak masking) and…

Alireza Salemi, Chang Zeng, Atharva Nijasure, Jui-Hui Chung, et al.

Published: May 2026
Upvotes: 117
Citations: 5

Read summary Original

39

arXiv.org

Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning

The paper introduces Skill1, a framework for training a single policy to co-evolve skill selection, utilization, and distillation in language model agents. Unlike prior methods that optimize these capabilities in isolation or with separate reward sources, Skill1 derives all learning signals from a single task-outcome reward by decomposing it into a…

Yaorui Shi, Yuxin Chen, Zhengxi Lu, Yuchun Miao, et al.

Published: May 2026
Upvotes: 116
Citations: 10

Read summary Original

40

NVIDIA

LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generation

LongLive-2.0 is an NVFP4-based parallel infrastructure for long video generation, co-designing training and inference. For training, it introduces Balanced SP, a sequence-parallel autoregressive (AR) training method that pairs clean-history and noisy-target temporal chunks on each GPU, enabling a natural teacher-forcing mask and SP-aware chunked VAE…

Yukang Chen, Luozhou Wang, Wei Huang, Shuai Yang, et al.

Published: May 2026
Upvotes: 116
Citations: 11

Read summary Original

41

arXiv.org

Rethinking Cross-Layer Information Routing in Diffusion Transformers

This paper investigates cross-layer information routing in Diffusion Transformers (DiTs), identifying three symptoms of standard residual connections: forward magnitude inflation, backward gradient decay, and block-wise redundancy. The authors propose Diffusion-Adaptive Routing (DAR), a drop-in replacement that uses learnable, timestep-adaptive,…

Chao Xu, Maohua Li, Qirui Li, Yixuan Xu, et al.

Published: May 2026
Upvotes: 113
Citations: 2

Read summary Original

42

arXiv.org

Lens: Rethinking Training Efficiency for Foundational Text-to-Image Models

Lens is a 3.8B-parameter text-to-image model that achieves performance competitive with larger state-of-the-art models while using significantly less training compute, requiring only about 19.3% of the compute used by Z-Image. Its efficiency stems from maximizing data information density per batch via the Lens-800M dataset of densely captioned image-text…

Dong Chen, Fangyun Wei, Ziyu Wan, Dongdong Chen, et al.

Published: May 2026
Upvotes: 111
Citations: 1

Read summary Original

43

arXiv.org

Stream-T1: Test-Time Scaling for Streaming Video Generation

Stream-T1 is a Test-Time Scaling (TTS) framework for streaming video generation, addressing the high costs and lack of temporal guidance in existing diffusion-based TTS methods. It leverages chunk-level synthesis and few denoising steps to reduce computational overhead. The framework comprises three components: Stream-Scaled Noise Propagation, which…

Yijing Tu, Shaojin Wu, Mengqi Huang, Wenchuan Wang, et al.

Published: May 2026
Upvotes: 109
Citations: 2

Read summary Original

44

arXiv.org

$π$-Bench: Evaluating Proactive Personal Assistant Agents in Long-Horizon Workflows

The paper introduces π-BENCH, a benchmark for evaluating proactive assistance in long-horizon personal assistant workflows. It comprises 100 multi-turn tasks across 5 domain-specific user personas (researcher, marketer, law trainee, pharmacist, financier), organized into 20-session episodes with cross-session dependencies. Each task begins with an…

Haoran Zhang, Luxin Xu, Zhilin Wang, Runquan Gui, et al.

Published: May 2026
Upvotes: 108
Citations: 1

Read summary Original

45

arXiv.org

OpenSearch-VL: An Open Recipe for Frontier Multimodal Search Agents

OpenSearch-VL is a fully open-source recipe for training multimodal deep search agents using agentic reinforcement learning. It addresses the lack of open high-quality training data, transparent trajectory synthesis, and detailed training recipes. The recipe includes a data curation pipeline using Wikipedia path sampling, fuzzy entity rewriting, and…

Shuang Chen, Kaituo Feng, Hangting Chen, Wenxuan Huang, et al.

Published: May 2026
Upvotes: 106
Citations: 10

Read summary Original

46

arXiv.org

WBench: A Comprehensive Multi-turn Benchmark for Interactive Video World Model Evaluation

WBENCH is a comprehensive multi-turn benchmark for evaluating interactive video world models across five dimensions: video quality, setting adherence, interaction adherence, consistency, and physics compliance. It contains 289 test cases and 1,058 interaction turns, covering diverse scenes, styles, subjects, and both first- and third-person perspectives.…

Kaining Ying, Hengrui Hu, Siyu Ren, Jiamu Li, et al.

Published: May 2026
Upvotes: 106
Citations: 7

Read summary Original

47

NVIDIA

AnyFlow: Any-Step Video Diffusion Model with On-Policy Flow Map Distillation

AnyFlow is a video diffusion distillation framework that enables any-step generation by learning flow-map transitions between arbitrary time pairs, unlike consistency models that degrade with more sampling steps. It uses a two-stage pipeline: forward flow map training (with interpolated timestep conditioning, guidance-fused training, and adaptive loss…

Yuchao Gu, Guian Fang, Yuxin Jiang, Weijia Mao, et al.

Published: May 2026
Upvotes: 105
Citations: 6

Read summary Original

48

arXiv.org

Flow-OPD: On-Policy Distillation for Flow Matching Models

Flow-OPD is a novel post-training framework for Flow Matching text-to-image models, addressing reward sparsity and gradient interference in multi-task alignment. It uses a two-stage strategy: first, domain-specialized teachers are trained via single-reward GRPO; then, a student model is initialized via a Flow-based Cold-Start (SFT or model merging) and…

Zhen Fang, Wenxuan Huang, Yu Zeng, Yiming Zhao, et al.

Published: May 2026
Upvotes: 102
Citations: 11

Read summary Original

49

Research paper

OCC-RAG: Optimal Cognitive Core for Faithful Question Answering

The paper introduces OCC-RAG, a family of small language models (SLMs) specialized for faithful, context-grounded question answering (QA). The models are mid-trained from Qwen3 base models (0.6B and 1.7B) on a synthetic corpus of over 3 million QA examples generated via a novel pipeline. The corpus includes single-hop, multi-hop (single- and…

Maksim Savkin, Mikhail Goncharov, Alexander Gambashidze, Alla Chepurova, et al.

Published: May 2026
Upvotes: 102
Citations: 0

Read summary Original

50

Research paper

ResearchClawBench: A Benchmark for End-to-End Autonomous Scientific Research

ResearchClawBench (RCBench) is a benchmark for evaluating end-to-end autonomous scientific research, comprising 40 tasks across 10 scientific domains (Astronomy, Chemistry, Earth Science, Energy, Information, Life, Material, Math, Neuroscience, Physics). Each task is derived from a real published paper, provides related literature and raw data, and hides…

Wanghan Xu, Shuo Li, Tianlin Ye, Qinglong Cao, et al.

Published: May 2026
Upvotes: 102
Citations: 3

Read summary Original

51

arXiv.org

Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps

RTPurbo is a method that transforms full-attention large language models into highly sparse models with minimal adaptation, requiring only a few hundred training steps. It is based on three observations: only a small subset of attention heads (retrieval heads) truly need full long-context processing; long-range retrieval is governed by a low-dimensional…

Yanke Zhou, Yiduo Li, Hanlin Tang, Maohua Li, et al.

Published: May 2026
Upvotes: 99
Citations: 1

Read summary Original

52

arXiv.org

Active Learners as Efficient PRP Rerankers

The paper reframes Pairwise Ranking Prompting (PRP) reranking as active learning from noisy pairwise comparisons, arguing that classical sorting algorithms are mismatched with noisy, order-sensitive LLM judgments. The authors propose active rankers, particularly the Mohajer algorithm, as drop-in replacements for sorting in call-constrained settings. They…

Jeremías Figueiredo Paschmann, Juan Kaplan, Francisco Nattero, Santiago Barron, et al.

Published: May 2026
Upvotes: 98
Citations: 0

Read summary Original

53

arXiv.org

Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation

Causal Forcing++ is a scalable pipeline for real-time interactive video generation that distills bidirectional diffusion models into few-step autoregressive (AR) students. It targets frame-wise autoregression with 1–2 sampling steps, a regime where existing initialization strategies fail: ODE distillation with a bidirectional teacher is architecturally…

Min Zhao, Hongzhou Zhu, Kaiwen Zheng, Zihan Zhou, et al.

Published: May 2026
Upvotes: 96
Citations: 11

Read summary Original

54

arXiv.org

Enhancing Train-Free Infinite-Frame Generation for Consistent Long Videos

The paper introduces MIGA, a training-free method for infinite-frame long video generation that builds on frame-level autoregressive frameworks like FIFO-Diffusion. MIGA addresses two key limitations: the training-inference gap and long-term consistency. It proposes a Two-Stage Training-Inference Alignment (TTA) mechanism that reduces the noise span of…

X. Feng, J. Zhu, M. Wu, C. Chen, et al.

Published: May 2026
Upvotes: 93
Citations: 1

Read summary Original

55

NVIDIA

Agent Explorative Policy Optimization for Multimodal Agentic Reasoning

The paper introduces AXPO (Agent eXplorative Policy Optimization) to address the Thinking-Acting Gap in multimodal agentic reasoning, where tool use is under-trained compared to thinking. Under GRPO, tool use occurs in only ~30% of rollouts, and tool-using subgroups are all-wrong on ~40% of questions, suppressing learning signals. AXPO fixes the thinking…

Minki Kang, Shizhe Diao, Ryo Hachiuma, Sung Ju Hwang, et al.

Published: May 2026
Upvotes: 93
Citations: 2

Read summary Original

56

NVIDIA

SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion Transformer

SANA-WM is a 2.6B-parameter open-source world model for generating one-minute, 720p videos with precise 6-DoF camera control. It uses a hybrid linear diffusion transformer combining frame-wise Gated DeltaNet (GDN) and softmax attention for efficient long-context modeling, a dual-branch camera control (UCPE and Plücker mixing), a two-stage generation…

Haoyi Zhu, Haozhe Liu, Yuyang Zhao, Tian Ye, et al.

Published: May 2026
Upvotes: 91
Citations: 13

Read summary Original

57

arXiv.org

Training Long-Context Vision-Language Models Effectively with Generalization Beyond 128K Context

This paper presents a systematic study of long-context continued pre-training (LongPT) for large vision-language models (LVLMs), extending a 7B model from 32K to 128K context. The authors find that long-document VQA is substantially more effective than OCR transcription for training. Key findings include: balanced sequence-length distributions outperform…

Zhaowei Wang, Lishu Luo, Haodong Duan, Weiwei Liu, et al.

Published: May 2026
Upvotes: 90
Citations: 0

Read summary Original

58

arXiv.org

OpenComputer: Verifiable Software Worlds for Computer-Use Agents

OpenComputer is a verifier-grounded framework for constructing verifiable software worlds for computer-use agents. It integrates four components: app-specific state verifiers, a self-evolving verification layer, a task-generation pipeline, and an evaluation harness. The framework covers 33 desktop applications and 1,000 finalized tasks. Experiments show…

Jinbiao Wei, Qianran Ma, Yilun Zhao, Xiao Zhou, et al.

Published: May 2026
Upvotes: 88
Citations: 6

Read summary Original

59

arXiv.org

ProRL: Effective Reinforcement Learning for Proactive Recommendation via Rectified Policy Gradient Estimation

The paper introduces ProRL, a reinforcement learning framework for proactive recommender systems (PRS) that guide user preferences toward target items via intermediate recommendation paths. The authors identify two deficiencies in standard policy gradient estimation for PRS: (1) a length shortcut, where positive-mean step-level rewards cause gradients to…

Hongru Hou, Tiehua Mei, Denghui Geng, Jinhui Huang, et al.

Published: May 2026
Upvotes: 88
Citations: 0

Read summary Original

60

arXiv.org

Continuous Latent Diffusion Language Model

Cola DLM is a hierarchical latent diffusion language model that decomposes text generation into global semantic modeling in a continuous latent space and local textual realization via a conditional decoder. It uses a Text VAE to map text to latent variables, a block-causal DiT to model the latent prior via flow matching, and a decoder for text generation.…

Hongcan Guo, Qinyu Zhao, Yian Zhao, Shen Nie, et al.

Published: May 2026
Upvotes: 87
Citations: 4

Read summary Original

61

ACM Transactions on Graphics

UniVidX: A Unified Multimodal Framework for Versatile Video Generation via Diffusion Priors

UniVidX is a unified multimodal framework for versatile video generation that repurposes video diffusion model (VDM) priors to handle diverse tasks within a single model. It addresses limitations of existing approaches that train separate models for fixed input-output mappings, ignoring cross-modal correlations. UniVidX introduces three key designs:…

Houyuan Chen, Hong Li, Xianghao Kong, Tianrui Zhu, et al.

Published: May 2026
Upvotes: 86
Citations: 0

Read summary Original

62

arXiv.org

IndusAgent: Reinforcing Open-Vocabulary Industrial Anomaly Detection with Agentic Tools

IndusAgent is a tool-augmented agentic framework for open-vocabulary industrial anomaly detection (IAD). It addresses limitations of multimodal large language models (MLLMs), such as domain-misaligned reasoning and structural hallucinations, by combining supervised fine-tuning (SFT) with reinforcement learning (RL). The framework constructs Indus-CoT, a…

Rongbin Tan, Fangfang Lin, Zhenlong Yuan, Min Qiu, et al.

Published: May 2026
Upvotes: 86
Citations: 0

Read summary Original

63

arXiv.org

Macaron-A2UI: A Model for Generative UI in Personal Agents

The paper introduces Macaron-A2UI, a model for Generative UI in personal agents, which dynamically synthesizes executable UI actions alongside natural language to improve interaction efficiency. The authors build a large-scale corpus of over 14,000 samples from four dialogue sources (MultiWOZ, SGD, ESConv, AnnoMI) using a hybrid rule-and-LLM pipeline with…

Fancy Kong, Congjie Zheng, Murphy Zhuang, Rio Yang, et al.

Published: May 2026
Upvotes: 84
Citations: 1

Read summary Original

64

arXiv.org

Soohak: A Mathematician-Curated Benchmark for Evaluating Research-level Math Capabilities of LLMs

SOOHAK is a new benchmark for evaluating research-level mathematical reasoning in LLMs, created by 64 mathematicians and comprising 340 Challenge and 99 Refusal problems, plus a 702-item SOOHAK-Mini subset. On the Challenge subset, frontier models like Gemini-3-Pro, GPT-5, and Claude-Opus-4.5 achieve Avg@3 scores of 30.39%, 26.37%, and 10.39% respectively,…

Guijin Son, Seungone Kim, Catherine Arnett, Hyunwoo Ko, et al.

Published: May 2026
Upvotes: 82
Citations: 4

Read summary Original

65

arXiv.org

Foundation Protocol: A Coordination Layer for Agentic Society

The Foundation Protocol (FP) is a graph-first coordination layer for an emerging human-AI society, designed to unify heterogeneous entities (agents, tools, humans, institutions, organizations) and support multi-party organization, event-based collaboration, economic primitives (metering, receipts, settlement), and first-class policy, provenance, and audit.…

Bang Liu, Yongfeng Gu, Jiayi Zhang, Zhaoyang Yu, et al.

Published: May 2026
Upvotes: 82
Citations: 0

Read summary Original

66

arXiv.org

EvalVerse: Pipeline-Aware and Expert-Calibrated Benchmarking for Professional Cinematic Video Generation

EvalVerse is a comprehensive evaluation framework for professional cinematic video generation that addresses the gap between basic prompt-following and true cinematic quality. It introduces a pipeline-aware taxonomy mirroring the filmmaking workflow (pre-production, production, post-production) with 3 stages, 7 aspects, 18 dimensions, 45 sub-dimensions,…

Songlin Yang, Haobin Zhong, Ruilin Zhang, Xiaotong Zhao, et al.

Published: May 2026
Upvotes: 82
Citations: 2

Read summary Original

67

arXiv.org

OmniRetrieval: Unified Retrieval across Heterogeneous Knowledge Sources

OmniRetrieval is a framework for unified retrieval across heterogeneous knowledge sources (unstructured text, relational databases, RDF knowledge graphs, and labeled property graphs). Instead of collapsing sources into a shared representation, it takes a natural-language query, selects relevant sources via a long-context LLM reading all source descriptors,…

Jinheon Baek, Soyeong Jeong, Sangwoo Park, Woongyeong Yeo, et al.

Published: May 2026
Upvotes: 81
Citations: 0

Read summary Original

68

Google DeepMind

RubricEM: Meta-RL with Rubric-guided Policy Decomposition beyond Verifiable Rewards

RubricEM is a reinforcement learning framework for training deep research agents on open-ended tasks where verifiable rewards are unavailable. It treats rubrics as a shared interface structuring policy execution, judge feedback, and agent memory. The framework decomposes trajectories into four rubric-guided stages (Plan, Research, Review, Answer) and uses…

Gaotang Li, Bhavana Dalvi Mishra, Zifeng Wang, Jun Yan, et al.

Published: May 2026
Upvotes: 79
Citations: 3

Read summary Original

69

NVIDIA

MemLens: Benchmarking Multimodal Long-Term Memory in Large Vision-Language Models

MEMLENS is a new benchmark for evaluating multimodal long-term memory in large vision-language models (LVLMs) and memory-augmented agents. It comprises 789 questions across five memory abilities (information extraction, multi-session reasoning, temporal reasoning, knowledge update, and answer refusal) at four context lengths (32K–256K tokens). An…

Xiyu Ren, Zhaowei Wang, Yiming Du, Zhongwei Xie, et al.

Published: May 2026
Upvotes: 79
Citations: 0

Read summary Original

70

arXiv.org

Lance: Unified Multimodal Modeling by Multi-Task Synergy

Lance is a lightweight native unified multimodal model from ByteDance that supports understanding, generation, and editing for both images and videos. It uses a dual-stream mixture-of-experts architecture on shared interleaved multimodal sequences, combining autoregressive language modeling for understanding with flow matching for generation. A…

Fengyi Fu, Mengqi Huang, Shaojin Wu, Yunsheng Jiang, et al.

Published: May 2026
Upvotes: 79
Citations: 3

Read summary Original

71

arXiv.org

EVA-Bench: A New End-to-end Framework for Evaluating Voice Agents

EVA-Bench is an end-to-end evaluation framework for voice agents that addresses two challenges: generating realistic simulated conversations and measuring quality across voice-specific failure modes. It uses bot-to-bot audio conversations with automatic simulation validation, and introduces two composite metrics: EVA-A (Accuracy) for task completion,…

Tara Bogavelli, Gabrielle Gauthier Melançon, Katrina Stankiewicz, Oluwanifemi Bamgbose, et al.

Published: May 2026
Upvotes: 76
Citations: 3

Read summary Original

72

arXiv.org

From Pixels to Words -- Towards Native One-Vision Models at Scale

NEO-ov is a native vision-language foundation model that unifies single-image, multi-image, video understanding, and spatial intelligence in a single monolithic backbone, eliminating external visual encoders, adapters, and post-hoc fusion. It uses a unified serialization scheme with spatiotemporal attention (THW-decoupled) and Native-RoPE to enable…

Haiwen Diao, Jiahao Wang, Penghao Wu, Yuhao Dong, et al.

Published: May 2026
Upvotes: 76
Citations: 1

Read summary Original

73

arXiv.org

A Matter of TASTE: Improving Coverage and Difficulty of Agent Benchmarks

The paper introduces TASTE (Task Synthesis from Tool Sequence Evolution), an automatic method for generating challenging agent benchmarks with broader tool-use coverage. TASTE reverses the traditional task construction process by first sampling diverse tool sequences using an Adaptive Contrastive n-gram model trained on LLM-judged validity signals, then…

Tomer Keren, Nitay Calderon, Asaf Yehudai, Yotam Perlitz, et al.

Published: May 2026
Upvotes: 75
Citations: 0

Read summary Original

74

arXiv.org

OpenSeeker-v2: Pushing the Limits of Search Agents with Informative and High-Difficulty Trajectories

OpenSeeker-v2 is a search agent developed by an academic team at Shanghai Jiao Tong University, demonstrating that a simple supervised fine-tuning (SFT) approach can rival resource-intensive industrial pipelines when trained on high-quality, high-difficulty trajectories. The authors introduce three data synthesis modifications: scaling the knowledge graph…

Yuwen Du, Rui Ye, Shuo Tang, Keduan Huang, et al.

Published: May 2026
Upvotes: 73
Citations: 6

Read summary Original

75

arXiv.org

SpatialBench: Is Your Spatial Foundation Model an All-Round Player?

SpatialBench is a new benchmark for evaluating spatial foundation models across diverse domains, input densities, and model paradigms. It includes 19 datasets, 546 scenes, 41 model variants, and 6 paradigms, using a deterministic multi-density sampling protocol (single, sparse, medium, dense). Key findings reveal that full-context attention models achieve…

Haosong Peng, Hao Li, Jiaqi Chen, Yuhao Pan, et al.

Published: May 2026
Upvotes: 72
Citations: 0

Read summary Original

76

arXiv.org

CollabVR: Collaborative Video Reasoning with Vision-Language and Video Generation Models

CollabVR is a closed-loop framework for video reasoning that couples a Vision-Language Model (VLM) with a Video Generation Model (VGM) at step-level granularity. It addresses two failure modes of VGMs: long-horizon drift and mid-clip simulation errors. The VLM plans the immediate next action, verifies the generated clip, and folds the diagnosis into the…

Joowon Kim, Seungho Shin, Joonhyung Park, Eunho Yang

Published: May 2026
Upvotes: 71
Citations: 1

Read summary Original

77

arXiv.org

World Action Models: The Next Frontier in Embodied AI

This survey introduces World Action Models (WAMs), a new paradigm in embodied AI that unifies predictive world modeling with action generation, targeting a joint distribution over future states and actions. WAMs are formally defined and distinguished from related concepts like Video Policies and Action World Models. The paper organizes existing methods…

Siyin Wang, Junhao Shi, Zhaoyang Fu, Xinzhe He, et al.

Published: May 2026
Upvotes: 71
Citations: 29

Read summary Original

78

arXiv.org

AI for Auto-Research: Roadmap & User Guide

This paper surveys AI-assisted research across the complete academic lifecycle, organized into four phases (Creation, Writing, Validation, Dissemination) and eight stages. It finds that AI excels at structured, retrieval-grounded tasks but remains unreliable for novel ideas, research-level experiments, and scientific judgment. Key findings include:…

Lingdong Kong, Xian Sun, Wei Chow, Linfeng Li, et al.

Published: May 2026
Upvotes: 71
Citations: 5

Read summary Original

79

Google DeepMind

LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling

The paper introduces AutoTTS, an environment-driven framework for automatically discovering test-time scaling (TTS) strategies for large language models, shifting the human role from hand-crafting heuristics to constructing discovery environments. The framework formulates width-depth TTS as controller synthesis over an offline replay environment built from…

Tong Zheng, Haolin Liu, Chengsong Huang, Huiwen Bao, et al.

Published: May 2026
Upvotes: 70
Citations: 0

Read summary Original

80

arXiv.org

Listwise Policy Optimization: Group-based RLVR as Target-Projection on the LLM Response Simplex

This paper introduces Listwise Policy Optimization (LPO), a framework for reinforcement learning with verifiable rewards (RLVR) in large language models. The authors show that existing group-based policy gradient methods (e.g., GRPO, Dr.GRPO, MaxRL) implicitly perform approximate target-projection on a response simplex, where the target is a…

Yun Qu, Qi Wang, Yixiu Mao, Heming Zou, et al.

Published: May 2026
Upvotes: 69
Citations: 1

Read summary Original

81

arXiv.org

Darwin Family: MRI-Trust-Weighted Evolutionary Merging for Training-Free Scaling of Language-Model Reasoning

The Darwin Family framework enables training-free evolutionary merging of large language models via gradient-free weight-space recombination, improving reasoning performance without additional training. It introduces a 14-dimensional adaptive merge genome, MRI-Trust Fusion (balancing diagnostic layer-importance signals with evolutionary search via a…

Taebong Kim, Youngsik Hong, Minsik Kim, Sunyoung Choi, et al.

Published: May 2026
Upvotes: 69
Citations: 0

Read summary Original

82

arXiv.org

Trust-Region Behavior Blending for On-Policy Distillation

The paper introduces Trust-Region behavior Blending (TRB), a warmup method for on-policy distillation (OPD) that addresses the issue of poor early student rollouts. TRB replaces the early rollout policy with a teacher-guided behavior policy that is constrained to stay within a student-centered KL trust region, while keeping the per-prefix reverse-KL OPD…

Daniil Plyusov, Alexey Gorbatovski, Alexey Malakhov, Nikita Balagansky, et al.

Published: May 2026
Upvotes: 69
Citations: 0

Read summary Original

83

arXiv.org

FashionChameleon: Towards Real-Time and Interactive Human-Garment Video Customization

FashionChameleon is a real-time and interactive framework for human-garment video customization, enabling users to switch garments during generation while preserving motion coherence. It uses three key techniques: a Teacher Model with In-Context Learning trained on single-garment data to implicitly handle garment switching; Streaming Distillation with…

Quanjian Song, Yefeng Shen, Mengting Chen, Hao Sun, et al.

Published: May 2026
Upvotes: 67
Citations: 3

Read summary Original

84

Together AI

OSCAR: Offline Spectral Covariance-Aware Rotation for 2-bit KV Cache Quantization

OSCAR is an INT2 KV-cache quantization method that uses attention-aware covariance structures, estimated offline, to derive fixed rotations and clipping thresholds, aligning quantization with downstream attention. It composes rotations as R = U · H_Had · P_br, where U is the eigenbasis of query-aware (Q^T Q) or score-aware (V^T S^T S V) covariance, H_Had…

Zhongzhu Zhou, Donglin Zhuang, Jisen Li, Ziyan Chen, et al.

Published: May 2026
Upvotes: 66
Citations: 2

Read summary Original

85

arXiv.org

MobileGym: A Verifiable and Highly Parallel Simulation Platform for Mobile GUI Agent Research

MOBILEGYM is a browser-hosted, lightweight simulation platform for mobile GUI agent research, targeting interaction fidelity without replicating proprietary backends. It enables verifiable outcome signals via deterministic state-based judging over structured JSON state, and scalable online RL through low-cost parallel rollouts. The environment state is…

Dingbang Wu, Rui Hao, Haiyang Wang, Shuzhe Wu, et al.

Published: May 2026
Upvotes: 66
Citations: 2

Read summary Original

86

arXiv.org

Do Enterprise Systems Need Learned World Models? The Importance of Context to Infer Dynamics

The paper investigates whether enterprise systems need learned world models, arguing that runtime discovery of configurable transition dynamics is more robust than offline training. It introduces CascadeBench, a benchmark for enterprise cascade prediction, and enterprise discovery agents that retrieve business rules at inference time. Experiments show…

Jishnu Sethumadhavan Nair, Patrice Bechard, Rishabh Maheshwary, Surajit Dasgupta, et al.

Published: May 2026
Upvotes: 65
Citations: 0

Read summary Original

87

arXiv.org

MemEye: A Visual-Centric Evaluation Framework for Multimodal Agent Memory

MemEye is a visual-centric evaluation framework for multimodal agent memory, addressing the gap in benchmarks that test whether agents preserve visual evidence for later reasoning. It introduces a two-axis taxonomy: visual evidence granularity (X-axis, from scene-level X1 to pixel-level X4) and memory reasoning depth (Y-axis, from atomic retrieval Y1 to…

Minghao Guo, Qingyue Jiao, Zeru Shi, Yihao Quan, et al.

Published: May 2026
Upvotes: 65
Citations: 6

Read summary Original

88

Research paper

SwanVoice: Expressive Long-Form Zero-Shot Speech Synthesis for Both Monologue and Dialogue

The paper introduces SwanVoice, a zero-shot text-to-speech (TTS) model for expressive long-form monologue and dialogue synthesis with 1–4 speakers. It addresses limitations of stitching monologue outputs for dialogue, which breaks acoustic consistency and affective continuity. The authors build SwanData-Speech, a data pipeline that processes 2.59 million…

Ruiqi Li, Yu Zhang, Changhao Pan, Ke Lei, et al.

Published: May 2026
Upvotes: 65
Citations: 2

Read summary Original

89

Research paper

Masking Stale Observations Helps Search Agents -- Until It Doesn't: A Regime Map and Its Mechanism

This paper investigates when masking stale observations in long-horizon search agents helps or hurts performance. The authors systematically vary backbone models (4B to 284B parameters) and retrievers (BM25, Qwen3-Emb-8B, AgentIR-4B) on offline (BrowseComp-Plus) and live-web (GAIA, xBench-DeepSearch, BrowseComp-ZH) benchmarks. They find that the accuracy…

Haoxiang Zhang, Qixin Xu, Zhuofeng Li, Lei Zhang, et al.

Published: May 2026
Upvotes: 65
Citations: 3

Read summary Original

90

arXiv.org

K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs

The paper introduces K12-KGraph, a curriculum-aligned knowledge graph built from official Chinese People's Education Press textbooks, covering mathematics, physics, chemistry, and biology across primary, middle, and high school. It includes nine node types and fourteen relation types, capturing both curriculum structure and visual grounding. From this…

Hao Liang, Qihan Lin, Zhaoyang Han, Xiaochen Ma, et al.

Published: May 2026
Upvotes: 64
Citations: 0

Read summary Original

91

arXiv.org

HyperEyes: Dual-Grained Efficiency-Aware Reinforcement Learning for Parallel Multimodal Search Agents

HyperEyes is a parallel multimodal search agent that addresses inefficiencies in sequential search agents for multi-entity queries. It introduces a Unified Grounded Search (UGS) action space, fusing visual grounding and retrieval into a single atomic action to enable concurrent search. Training uses a two-stage approach: a Parallel-Amenable Data Synthesis…

Guankai Li, Jiabin Chen, Yi Xu, Xichen Zhang, et al.

Published: May 2026
Upvotes: 63
Citations: 4

Read summary Original

92

arXiv.org

Representation Forcing for Bottleneck-Free Unified Multimodal Models

The paper introduces Representation Forcing (RF), a technique for unified multimodal models (UMMs) that eliminates the need for a separately pretrained VAE in image generation. RF trains the decoder to autoregressively predict discrete visual representation tokens, derived from the model's own understanding encoder via online vector quantization, before…

Yuqing Wang, Zhijie Lin, Ceyuan Yang, Yang Zhao, et al.

Published: May 2026
Upvotes: 63
Citations: 3

Read summary Original

93

Qwen

Qwen-Image-VAE-2.0 Technical Report

Qwen-Image-VAE-2.0 is a suite of high-compression image VAEs (f16 and f32) designed to overcome the trade-off between compression ratio, reconstruction fidelity, and diffusability. The architecture uses Global Skip Connections (GSC) to preserve fine details, expanded latent channels, and an attention-free, asymmetric encoder-decoder backbone for…

Zekai Zhang, Deqing Li, Kuan Cao, Yujia Wu, et al.

Published: May 2026
Upvotes: 62
Citations: 1

Read summary Original

94

arXiv.org

CollectionLoRA: Collecting 50 Effects in 1 LoRA via Multi-Teacher On-Policy Distillation

CollectionLoRA is a multi-teacher on-policy distillation framework that consolidates up to 50 diverse visual effects and few-step generation capabilities into a single LoRA, addressing storage overhead, routing latency, and parameter conflicts in conventional multi-LoRA pipelines. The method introduces three key components: Probabilistic Dual-Stream…

Fangtai Wu, Hailong Guo, Shijie Huang, Jiayi Song, et al.

Published: May 2026
Upvotes: 62
Citations: 2

Read summary Original

95

arXiv.org

Self-Improving Language Models with Bidirectional Evolutionary Search

The paper introduces Bidirectional Evolutionary Search (BES), a search framework for self-improving language models and agents. BES addresses two limitations of existing methods (best-of-N and tree search): sparse verification signals and confinement to the model's own distribution. It couples forward search, which uses expansion and four evolution…

Guowei Xu, Zhenting Qi, Huangyuan Su, Weirui Ye, et al.

Published: May 2026
Upvotes: 62
Citations: 1

Read summary Original

96

arXiv.org

Learning to Foresee: Unveiling the Unlocking Efficiency of On-Policy Distillation

This paper investigates why on-policy distillation (OPD) is more efficient than reinforcement learning (RL) for post-training large language models. The authors identify two properties of OPD's parameter updates: Functional Redundancy Avoidance, where OPD suppresses updates in low-utility modules (e.g., embeddings, peripheral layers) and concentrates on…

Yuchen Cai, Ding Cao, Liang Lin, Chunxi Luo, et al.

Published: May 2026
Upvotes: 61
Citations: 4

Read summary Original

97

arXiv.org

ACC: Compiling Agent Trajectories for Long-Context Training

The paper introduces Agent Context Compilation (ACC), a method that converts multi-turn agent trajectories (from search, software engineering, and SQL agents) into long-context QA training pairs. Standard agent SFT masks tool responses, creating a supervision blind spot where evidence scattered across turns is unused. ACC compiles tool responses and…

Qisheng Su, Zhen Fang, Shiting Huang, Yu Zeng, et al.

Published: May 2026
Upvotes: 61
Citations: 0

Read summary Original

98

arXiv.org

MCP-Cosmos: World Model-Augmented Agents for Complex Task Execution in MCP Environments

MCP-Cosmos is a framework that integrates generative World Models (WM) into the Model Context Protocol (MCP) ecosystem to enable predictive task automation. It uses a 'Bring Your Own World Model' (BYOWM) strategy, allowing agents to simulate state transitions and refine plans in a latent space before execution. The framework was evaluated on 24 MCP-Bench…

Giridhar Ganapavarapu, Dhaval Patel

Published: May 2026
Upvotes: 60
Citations: 1

Read summary Original

99

arXiv.org

SciAtlas: A Large-Scale Knowledge Graph for Automated Scientific Research

SciAtlas is a large-scale, multi-disciplinary knowledge graph designed to support automated scientific research by organizing over 43 million papers from 26 disciplines into a structured topological network. It contains 157 million entities (including papers, authors, institutions, keywords, topics, and fields) and 3 billion triplets, with 9 entity types…

Shuofei Qiao, Yunxiang Wei, Jiazheng Fan, Bin Wu, et al.

Published: May 2026
Upvotes: 60
Citations: 0

Read summary Original

100

NVIDIA

Why Far Looks Up: Probing Spatial Representation in Vision-Language Models

The paper investigates whether vision-language models (VLMs) achieve spatial reasoning through structured 3D understanding or by exploiting statistical shortcuts in natural images. The authors introduce a representation-level analysis framework using contrastive pairs to measure how spatial axes (horizontal, vertical, depth) are organized in VLM…

Cheolhong Min, Jaeyun Jung, Daeun Lee, Hyeonseong Jeon, et al.

Published: May 2026
Upvotes: 60
Citations: 1

Read summary Original

101

Research paper

Formalizing Latent Thoughts: Four Axioms of Thought Representation in LLMs

This paper introduces an axiomatic evaluation framework for latent thought representations in LLMs, defining four functional axioms: Causality, Minimality, Separability, and Stability. Each axiom is quantified by a metric computed directly on the representation, independent of downstream task accuracy. The authors audit five open-weight LLMs (Llama-3.1 8B,…

Fahd Seddik, Fatemeh Fard

Published: May 2026
Upvotes: 60
Citations: 0

Read summary Original

102

arXiv.org

GoLongRL: Capability-Oriented Long Context Reinforcement Learning with Multitask Alignment

GoLongRL is a fully open-source, capability-oriented post-training recipe for long-context reinforcement learning with verifiable rewards (RLVR). It introduces a dataset of 22,965 samples spanning 9 task types, each paired with its natural evaluation metric (e.g., EM, F1, NDCG), and a four-phase construction pipeline. The dataset combines curated…

Minxuan Lv, Tiehua Mei, Tanlong Du, Junmin Chen, et al.

Published: May 2026
Upvotes: 59
Citations: 2

Read summary Original

103

arXiv.org

minWM: A Full-Stack Open-Source Framework for Real-Time Interactive Video World Models

minWM is a full-stack open-source framework for converting bidirectional text-to-video (T2V) or text-and-image-to-video (TI2V) diffusion foundation models into camera-controllable, few-step autoregressive (AR) world models for real-time interaction. The pipeline has two phases: first, fine-tuning the bidirectional model with camera control via PRoPE…

Min Zhao, Hongzhou Zhu, Bokai Yan, Zihan Zhou, et al.

Published: May 2026
Upvotes: 59
Citations: 6

Read summary Original

104

arXiv.org

Mellum2 Technical Report

Mellum 2 is an open-weight 12B-parameter Mixture-of-Experts (MoE) language model with 2.5B active parameters per token, designed for software engineering tasks. It uses 64 experts with 8 active, Grouped-Query Attention with 4 KV heads, Sliding Window Attention on 3 of every 4 layers, and a Multi-Token Prediction head for speculative decoding. Pre-training…

Marko Kojic, Ivan Bondyrev, Aral de Moor, Joseph Shtok, et al.

Published: May 2026
Upvotes: 59
Citations: 0

Read summary Original

105

arXiv.org

MiA-Signature: Approximating Global Activation for Long-Context Understanding

The paper introduces Mindscape Activation Signature (MiA-Signature), a compact representation of the global activation pattern induced by a query over a semantic memory space, inspired by cognitive science theories of global ignition and partial access. It is constructed via submodular selection of high-level concepts (session summaries) that cover the…

Yuqing Li, Jiangnan Li, Mo Yu, Zheng Lin, et al.

Published: May 2026
Upvotes: 57
Citations: 1

Read summary Original

106

arXiv.org

A Survey of Large Audio Language Models: Generalization, Trustworthiness, and Outlook

This survey provides a comprehensive analysis of Large Audio Language Models (LALMs), focusing on their generalization, trustworthiness, and future outlook. It examines the endogenous mechanisms of LALMs, including architectural foundations, representational paradigms, training and alignment strategies, and emergent reasoning mechanisms. The survey…

Kaiwen Luo, Zhenhong Zhou, Leyan Wang, Liang Lin, et al.

Published: May 2026
Upvotes: 56
Citations: 4

Read summary Original

107

arXiv.org

YoCausal: How Far is Video Generation from World Model? A Causality Perspective

YoCausal is a two-level benchmark for evaluating causal cognition in video diffusion models (VDMs), inspired by the Violation of Expectation paradigm. It uses temporally reversed real-world videos as counterfactual samples, avoiding synthetic data and the sim-to-real gap. Level 1 introduces the Reverse Surprise Index (RSI), measuring arrow-of-time…

You-Zhe Xie, Yu-Hsuan Li, Jie-Ying Lee, Kaipeng Zhang, et al.

Published: May 2026
Upvotes: 56
Citations: 1

Read summary Original

108

arXiv.org

HumanNet: Scaling Human-centric Video Learning to One Million Hours

HumanNet is a one-million-hour human-centric video corpus designed to scale embodied learning by capturing how humans interact with the physical world. It includes both first-person and third-person perspectives, covering fine-grained activities, human-object interactions, tool use, and long-horizon behaviors across diverse environments. The dataset…

Yufan Deng, Daquan Zhou

Published: May 2026
Upvotes: 55
Citations: 6

Read summary Original

109

arXiv.org

Auditing Agent Harness Safety

This paper introduces HarnessAudit, a framework for auditing LLM agent execution harnesses, and HarnessAudit-Bench, a benchmark of 210 tasks across 8 domains. The framework evaluates full execution trajectories on three layers: boundary compliance (tool, resource, and information-flow violations), execution fidelity (action validity and task completion),…

Chengzhi Liu, Yichen Guo, Yepeng Liu, Yuzhe Yang, et al.

Published: May 2026
Upvotes: 55
Citations: 1

Read summary Original

110

arXiv.org

DexJoCo: A Benchmark and Toolkit for Task-Oriented Dexterous Manipulation on MuJoCo

DexJoCo is a benchmark and toolkit for task-oriented dexterous manipulation on MuJoCo, featuring 11 functionally grounded tasks that evaluate tool-use, bimanual coordination, long-horizon execution, and reasoning. It includes a low-cost teleoperation system using Rokoko gloves and HTC Vive trackers, with a retargeting module (GeoRT) to map human hand…

Hanwen Wang, Weizhi Zhao, Xiangyu Wang, Siyuan Huang, et al.

Published: May 2026
Upvotes: 55
Citations: 5

Read summary Original

111

arXiv.org

CHI-Bench: Can AI Agents Automate End-to-End, Long-Horizon, Policy-Rich Healthcare Workflows?

χ-Bench is a benchmark for evaluating AI agents on long-horizon, policy-rich healthcare workflows across three domains: provider prior authorization, payer utilization management, and care management. It uses a high-fidelity simulator of 20 healthcare apps exposed via 87 MCP tools, guided by a 1,279-document managed-care operations handbook. Across 30…

Haolin Chen, Deon Metelski, Leon Qi, Tao Xia, et al.

Published: May 2026
Upvotes: 55
Citations: 4

Read summary Original

112

arXiv.org

PhysX-Omni: Unified Simulation-Ready Physical 3D Generation for Rigid, Deformable, and Articulated Objects

PhysX-Omni is a unified framework for generating simulation-ready physical 3D assets covering rigid, deformable, and articulated objects. It introduces a novel template-based run-length encoding (RLE) geometry representation for Vision-Language Models, which directly encodes high-resolution 3D structures without special tokens or segmentation modules,…

Ziang Cao, Yinghao Liu, Haitian Li, Runmao Yao, et al.

Published: May 2026
Upvotes: 55
Citations: 1

Read summary Original

113

Research paper

DataPrep-Bench: Benchmarking LLMs as Training Data Preparators

DataPrep-Bench is the first unified benchmark for evaluating LLMs as training data preparators, covering two complementary capabilities: data construction (transforming raw sources into supervised training data) and data quality evaluation (predicting downstream training utility of candidate datasets). The benchmark is grounded in downstream performance…

Hao Liang, Qifeng Cai, Yibo Lin, Jianzhuo Du, et al.

Published: May 2026
Upvotes: 55
Citations: 0

Read summary Original

114

arXiv.org

From Activation to Causality: Discovery of Causal Visual Representations in the Human Brain

BrainCause is an automated framework for causally discovering and validating visual concept representations in the human brain using fMRI. It addresses the limitation of activation-based methods, which often identify false positives driven by correlated visual or semantic cues. BrainCause constructs targeted stimulus sets with positive images,…

Yuval Golbari, Navve Wasserman, Matias Cosarinsky, Roman Beliy, et al.

Published: May 2026
Upvotes: 54
Citations: 0

Read summary Original

115

MiniMax

The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence

The MiniMax-M2 series is a family of Mixture-of-Experts language models designed to achieve frontier performance with minimal activated parameters. The flagship M2 has 229.9B total parameters but only 9.8B activated per token, using a 62-layer decoder-only Transformer with 256 fine-grained experts, sigmoid gating, full multi-head attention with GQA, a…

Aili Chen, Aonian Li, Baichuan Zhou, Bangwei Gong, et al.

Published: May 2026
Upvotes: 41
Citations: 23

Read summary Original