Monthly collection

September 2025

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

Papers: 79
Citations: 4.4K
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

79 papers published in September 2025

01

arXiv.org

Sharing is Caring: Efficient LM Post-Training with Collective RL Experience Sharing

The paper introduces Swarm sAmpling Policy Optimization (SAPO), a fully decentralized and asynchronous reinforcement learning (RL) post-training algorithm for language models (LMs). SAPO enables heterogeneous compute nodes to train their own policies while sharing decoded rollouts with the swarm, avoiding synchronization bottlenecks and hardware…

Jeffrey Amico, Gabriel Passamani Andrade, John Donaghy, Ben Fielding, et al.

Published: Sep 2025
Upvotes: 665
Citations: 3

Read summary Original

02

arXiv.org

The Dragon Hatchling: The Missing Link between the Transformer and Models of the Brain

The paper introduces Dragon Hatchling (BDH), a new Large Language Model architecture based on a scale-free, biologically inspired network of n locally-interacting neuron particles. BDH couples strong theoretical foundations and inherent interpretability without sacrificing Transformer-like performance. It is a practical, performant attention-based state…

Adrian Kosowski, Przemysław Uznański, Jan Chorowski, Zuzanna Stamirowska, et al.

Published: Sep 2025
Upvotes: 551
Citations: 5

Read summary Original

03

AAAI Conference on Artificial Intelligence

VLA-Adapter: An Effective Paradigm for Tiny-Scale Vision-Language-Action Model

VLA-Adapter is a novel paradigm for Vision-Language-Action (VLA) models that reduces reliance on large-scale VLMs and extensive pre-training. It systematically analyzes which vision-language (VL) conditions are essential for bridging perception to action, finding that middle-layer raw features and deep-layer ActionQuery features are most effective, and…

Yihao Wang, Pengxiang Ding, Lingxiao Li, Can Cui, et al.

Published: Sep 2025
Upvotes: 259
Citations: 129

Read summary Original

04

arXiv.org

The Landscape of Agentic Reinforcement Learning for LLMs: A Survey

This survey formalizes Agentic Reinforcement Learning (Agentic RL), a paradigm where LLMs are treated as learnable policies in sequential decision-making loops, contrasting it with conventional single-step LLM RL. It proposes a twofold taxonomy: one based on core agentic capabilities (planning, tool use, memory, reasoning, self-improvement, perception) and…

Guibin Zhang, Hejia Geng, Xiaohang Yu, Zhenfei Yin, et al.

Published: Sep 2025
Upvotes: 239
Citations: 182

Read summary Original

05

Conference on Empirical Methods in Natural Language Processing

Drivel-ology: Challenging LLMs with Interpreting Nonsense with Depth

The paper introduces Drivelology, a linguistic phenomenon of syntactically coherent but pragmatically paradoxical utterances that encode implicit meaning requiring contextual inference, moral reasoning, or emotional interpretation. The authors construct DRIVELHUB, a benchmark dataset of over 1,200 curated examples across six languages (English, Mandarin,…

Yang Wang, Chenghao Xiao, Chia-Yi Hsiao, Zi Yan Chang, et al.

Published: Sep 2025
Upvotes: 213
Citations: 3

Read summary Original

06

OpenAI

Why Language Models Hallucinate

The paper argues that language model hallucinations arise from statistical pressures during pretraining and persist due to misaligned evaluation metrics. The authors formalize hallucinations as errors in binary classification, showing that even with error-free training data, the cross-entropy objective leads to errors. They introduce the Is-It-Valid (IIV)…

Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, Edwin Zhang

Published: Sep 2025
Upvotes: 200
Citations: 282

Read summary Original

07

arXiv.org

A Survey of Reinforcement Learning for Large Reasoning Models

This survey reviews recent advances in Reinforcement Learning (RL) for Large Reasoning Models (LRMs), focusing on how RL transforms LLMs into LRMs, particularly since DeepSeek-R1. It covers foundational components: reward design (verifiable, generative, dense, unsupervised, and shaping), policy optimization (critic-based, critic-free, off-policy, and…

Kaiyan Zhang, Yuxin Zuo, Bingxiang He, Youbang Sun, et al.

Published: Sep 2025
Upvotes: 193
Citations: 161

Read summary Original

08

NVIDIA

LongLive: Real-time Interactive Long Video Generation

LONGLIVE is a frame-level autoregressive (AR) framework for real-time, interactive long video generation, addressing efficiency and quality challenges in diffusion and AR models. It introduces KV-recache to refresh cached states with new prompts for smooth, adherent prompt switches; streaming long tuning to enable train-long-test-long alignment; and…

Shuai Yang, Wei Huang, Ruihang Chu, Yicheng Xiao, et al.

Published: Sep 2025
Upvotes: 190
Citations: 189

Read summary Original

09

arXiv.org

MCPMark: A Benchmark for Stress-Testing Realistic and Comprehensive MCP Use

MCPMark is a benchmark designed to evaluate LLM agents' use of the Model Context Protocol (MCP) in realistic, comprehensive workflows. It comprises 127 tasks across five MCP servers (Filesystem, Notion, Playwright, GitHub, PostgreSQL), each with curated initial states and programmatic verification scripts. Tasks require diverse CRUD operations and average…

Zijian Wu, Xiangyan Liu, Xinyuan Zhang, Lingjun Chen, et al.

Published: Sep 2025
Upvotes: 180
Citations: 25

Read summary Original

10

arXiv.org

MinerU2.5: A Decoupled Vision-Language Model for Efficient High-Resolution Document Parsing

MinerU2.5 is a 1.2B-parameter vision-language model for document parsing that achieves state-of-the-art accuracy with high efficiency. It uses a decoupled, two-stage strategy: first, global layout analysis on a downsampled 1036x1036 image; second, targeted content recognition on native-resolution crops guided by the layout. The model uses a 675M NaViT…

Junbo Niu, Zheng Liu, Zhuangcheng Gu, Bin Wang, et al.

Published: Sep 2025
Upvotes: 177
Citations: 86

Read summary Original

11

Qwen

Qwen3-Omni Technical Report

Qwen3-Omni is a single multimodal model that achieves state-of-the-art performance across text, image, audio, and video without degradation relative to single-modal counterparts. It matches the performance of same-sized Qwen single-modal models and excels on audio tasks, achieving open-source SOTA on 32 of 36 audio benchmarks and overall SOTA on 22,…

Jin Xu, Zhifang Guo, Hangrui Hu, Yunfei Chu, et al.

Published: Sep 2025
Upvotes: 154
Citations: 405

Read summary Original

12

arXiv.org

Reverse-Engineered Reasoning for Open-Ended Generation

The paper introduces REverse-Engineered Reasoning (REER), a new paradigm for instilling deep reasoning in LLMs for open-ended, non-verifiable tasks like creative writing, where reinforcement learning (RL) and instruction distillation fail due to lack of clear rewards and high costs. REER works backwards from known good solutions, using a gradient-free…

Haozhe Wang, Haoran Que, Qixin Xu, Minghao Liu, et al.

Published: Sep 2025
Upvotes: 151
Citations: 33

Read summary Original

13

arXiv.org

DeepSearch: Overcome the Bottleneck of Reinforcement Learning with Verifiable Rewards via Monte Carlo Tree Search

DeepSearch is a framework that integrates Monte Carlo Tree Search (MCTS) directly into reinforcement learning with verifiable rewards (RLVR) training to overcome exploration bottlenecks that cause performance plateaus. Unlike methods that use tree search only at inference, DeepSearch embeds structured search into the training loop, enabling systematic…

Fang Wu, Weihao Xuan, Heli Qi, Ximing Lu, et al.

Published: Sep 2025
Upvotes: 147
Citations: 10

Read summary Original

14

arXiv.org

Vision-Zero: Scalable VLM Self-Improvement via Strategic Gamified Self-Play

Vision-Zero is a label-free, domain-agnostic multi-agent self-play framework for self-evolving vision-language models (VLMs) through competitive visual games generated from arbitrary images. It trains VLMs in a 'Who Is the Spy?'-style game where civilians see an image and the spy sees a blank input, requiring strategic reasoning and communication. The…

Qinsi Wang, Bo Liu, Tianyi Zhou, Jing Shi, et al.

Published: Sep 2025
Upvotes: 142
Citations: 33

Read summary Original

15

arXiv.org

EPO: Entropy-regularized Policy Optimization for LLM Agents Reinforcement Learning

The paper identifies the exploration-exploitation cascade failure in multi-turn LLM agent reinforcement learning, where shared policy parameters across turns and sparse terminal rewards cause severe entropy oscillations that destabilize training. To address this, the authors propose Entropy-regularized Policy Optimization (EPO), which combines…

Wujiang Xu, Wentian Zhao, Zhenting Wang, Yu-Jhe Li, et al.

Published: Sep 2025
Upvotes: 137
Citations: 10

Read summary Original

16

arXiv.org

Baseer: A Vision-Language Model for Arabic Document-to-Markdown OCR

The paper introduces Baseer, a vision-language model fine-tuned for Arabic document OCR, and Misraj-DocOCR, a benchmark for evaluation. Baseer is built on Qwen2.5-VL-3B-Instruct and trained on 500,000 image-text pairs (300k synthetic, 200k real-world) using a decoder-only fine-tuning strategy that freezes the vision encoder. The authors also corrected the…

Khalil Hennara, Muhammad Hreden, Mohamed Motasim Hamed, Ahmad Bastati, et al.

Published: Sep 2025
Upvotes: 135
Citations: 2

Read summary Original

17

arXiv.org

HuMo: Human-Centric Video Generation via Collaborative Multi-Modal Conditioning

HuMo is a unified framework for Human-Centric Video Generation (HCVG) that enables collaborative control from text, reference images, and audio. It addresses two challenges: data scarcity and difficulty in coordinating sub-tasks of subject preservation and audio-visual sync. To overcome data scarcity, HuMo constructs a high-quality dataset with paired…

Liyang Chen, Tianxiang Ma, Jiawei Liu, Bingchuan Li, et al.

Published: Sep 2025
Upvotes: 130
Citations: 45

Read summary Original

18

arXiv.org

RPG: A Repository Planning Graph for Unified and Scalable Codebase Generation

The paper introduces the Repository Planning Graph (RPG), a structured representation that unifies proposal-level and implementation-level planning for generating complete software repositories from high-level specifications. RPG encodes capabilities, file structures, data flows, and functions as nodes and edges, replacing ambiguous natural language plans.…

Jane Luo, Xin Zhang, Steven Liu, Jie Wu, et al.

Published: Sep 2025
Upvotes: 129
Citations: 12

Read summary Original

19

arXiv.org

UI-TARS-2 Technical Report: Advancing GUI Agent with Multi-Turn Reinforcement Learning

UI-TARS-2 is a native GUI-centered agent model developed by ByteDance Seed, designed to handle both structured computer-use tasks and dynamic game environments. It addresses challenges in data scalability, multi-turn reinforcement learning (RL), GUI-only operation limits, and environment stability through a systematic methodology: a data flywheel for…

Haoming Wang, Haoyang Zou, Huatong Song, Jiazhan Feng, et al.

Published: Sep 2025
Upvotes: 128
Citations: 163

Read summary Original

20

arXiv.org

VCRL: Variance-based Curriculum Reinforcement Learning for Large Language Models

The paper introduces VCRL, a curriculum reinforcement learning framework for large language models (LLMs) that dynamically adjusts training sample difficulty based on the variance of group rewards in rollout-based RL (e.g., GRPO). The authors observe that samples with moderate difficulty yield higher reward variance, while too-easy or too-hard samples…

Guochao Jiang, Wenfeng Feng, Guofeng Quan, Chuzhan Hao, et al.

Published: Sep 2025
Upvotes: 122
Citations: 27

Read summary Original

21

arXiv.org

FlowRL: Matching Reward Distributions for LLM Reasoning

FlowRL is a policy optimization algorithm for large language model (LLM) reasoning that shifts from reward maximization to reward distribution matching. It uses a learnable partition function to normalize scalar rewards into a target distribution and minimizes the reverse KL divergence between the policy and this distribution, which is equivalent to a…

Xuekai Zhu, Daixuan Cheng, Dinghuai Zhang, Hengli Li, et al.

Published: Sep 2025
Upvotes: 119
Citations: 37

Read summary Original

22

Research paper

Quantile Advantage Estimation: Stabilizing RLVR for LLM Reasoning

The paper introduces Quantile Advantage Estimation (QAE) to stabilize Reinforcement Learning with Verifiable Rewards (RLVR) for LLM reasoning. It identifies that mean-baseline methods like GRPO and DAPO cause entropy collapse or explosion, harming performance. QAE replaces the mean reward baseline with a group-wise K-quantile, creating a two-regime gate:…

Junkang Wu, Kexin Huang, Jiancan Wu, An Zhang, et al.

Published: Sep 2025
Upvotes: 119
Citations: 10

Read summary Original

23

arXiv.org

SLA: Beyond Sparsity in Diffusion Transformers via Fine-Tunable Sparse-Linear Attention

The paper introduces SLA (Sparse-Linear Attention), a trainable attention method for Diffusion Transformers (DiTs) that combines sparse and linear attention to reduce computational cost. The authors observe that attention weights can be decomposed into a small fraction of large weights with high rank and a large fraction of low-rank weights. SLA classifies…

Jintao Zhang, Haoxu Wang, Kai Jiang, Shuo Yang, et al.

Published: Sep 2025
Upvotes: 119
Citations: 44

Read summary Original

24

Qwen

Scaling Agents via Continual Pre-training

The paper introduces Agentic Continual Pre-training (Agentic CPT), a new training stage between pre-training and post-training, to build agentic foundation models for deep research agents. The authors argue that post-training on general-purpose models creates optimization conflicts, as models must simultaneously learn agentic behaviors and align to expert…

Liangcai Su, Zhen Zhang, Guangyu Li, Zhuo Chen, et al.

Published: Sep 2025
Upvotes: 118
Citations: 41

Read summary Original

25

arXiv.org

ScaleCUA: Scaling Open-Source Computer Use Agents with Cross-Platform Data

ScaleCUA introduces a large-scale, cross-platform dataset and model family for computer use agents (CUAs), addressing data scarcity via a dual-loop pipeline combining automated agents and human experts across six platforms (Windows, macOS, Linux, Android, iOS, Web). The dataset includes 471K GUI understanding examples, 17.1M grounding annotations, and 19K…

Zhaoyang Liu, Jingjing Xie, Zichen Ding, Zehao Li, et al.

Published: Sep 2025
Upvotes: 111
Citations: 46

Read summary Original

26

arXiv.org

OmniWorld: A Multi-Domain and Multi-Modal Dataset for 4D World Modeling

OmniWorld is a large-scale, multi-domain, multi-modal dataset for 4D world modeling, introduced by Shanghai AI Lab and ZJU. It comprises a self-collected OmniWorld-Game synthetic dataset (96K clips, 18.5M frames, 214+ hours) and curated public datasets from robot, human, and internet domains, totaling over 600K sequences and 300M frames. OmniWorld provides…

Yang Zhou, Yifan Wang, Jianjun Zhou, Wenzheng Chang, et al.

Published: Sep 2025
Upvotes: 107
Citations: 43

Read summary Original

27

arXiv.org

WebWeaver: Structuring Web-Scale Evidence with Dynamic Outlines for Open-Ended Deep Research

WebWeaver is a dual-agent framework for open-ended deep research (OEDR), addressing limitations of static pipelines and monolithic generation. It comprises a planner that iteratively interleaves evidence acquisition with outline optimization, producing a citation-grounded outline linked to a memory bank, and a writer that performs hierarchical,…

Zijian Li, Xin Guan, Bo Zhang, Shen Huang, et al.

Published: Sep 2025
Upvotes: 107
Citations: 39

Read summary Original

28

arXiv.org

Parallel-R1: Towards Parallel Thinking via Reinforcement Learning

Parallel-R1 is the first reinforcement learning (RL) framework to train large language models (LLMs) in parallel thinking for general mathematical reasoning. It uses a progressive curriculum: supervised fine-tuning (SFT) on easy GSM8K problems to teach the parallel thinking format, followed by RL on harder DAPO problems to generalize the skill. The…

Tong Zheng, Hongming Zhang, Wenhao Yu, Xiaoyang Wang, et al.

Published: Sep 2025
Upvotes: 106
Citations: 49

Read summary Original

29

arXiv.org

LIMI: Less is More for Agency

The paper introduces LIMI (Less Is More for Intelligent Agency), which challenges the assumption that more training data yields better agentic AI. LIMI demonstrates that sophisticated agentic intelligence can emerge from minimal, strategically curated demonstrations. Using only 78 carefully designed training samples focused on vibe coding and research…

Yang Xiao, Mohan Jiang, Jie Sun, Keyu Li, et al.

Published: Sep 2025
Upvotes: 104
Citations: 17

Read summary Original

30

arXiv.org

MMR1: Enhancing Multimodal Reasoning with Variance-Aware Sampling and Open Resources

This paper introduces MMR1, a framework for enhancing multimodal reasoning in large models. It addresses two key limitations: the lack of open, large-scale, high-quality long chain-of-thought (CoT) data, and the instability of reinforcement learning (RL) algorithms, specifically gradient vanishing in Group Relative Policy Optimization (GRPO) when reward…

Sicong Leng, Jing Wang, Jiaxi Li, Hao Zhang, et al.

Published: Sep 2025
Upvotes: 104
Citations: 35

Read summary Original

31

Google DeepMind

Video models are zero-shot learners and reasoners

This paper investigates whether generative video models, like Veo 3, can act as zero-shot learners and reasoners for general-purpose vision tasks, similar to how LLMs transformed NLP. The authors analyzed 18,384 generated videos across 62 qualitative and 7 quantitative tasks, finding that Veo 3 can solve tasks it wasn't explicitly trained for, including…

Thaddäus Wiedemer, Yuxuan Li, Paul Vicol, Shixiang Shane Gu, et al.

Published: Sep 2025
Upvotes: 101
Citations: 194

Read summary Original

32

arXiv.org

SciReasoner: Laying the Scientific Reasoning Ground Across Disciplines

SciReasoner is a scientific reasoning foundation model that aligns natural language with heterogeneous scientific representations. It is pretrained on a 206B-token corpus (scientific text, pure sequences, sequence-text pairs) and post-trained via SFT on 40M instructions, annealed cold-start bootstrapping for chain-of-thought, and reinforcement learning…

Yizhou Wang, Chen Tang, Han Deng, Jiabei Xiao, et al.

Published: Sep 2025
Upvotes: 101
Citations: 9

Read summary Original

33

arXiv.org

From Editor to Dense Geometry Estimator

FE2E is a framework that adapts a pre-trained image editing model, Step1X-Edit, for monocular dense geometry prediction (depth and normal estimation). The authors argue that editing models, unlike text-to-image generators, possess inherent structural priors that make them more suitable for image-to-image tasks. They introduce three key adaptations: a…

JiYuan Wang, Chunyu Lin, Lei Sun, Rongying Liu, et al.

Published: Sep 2025
Upvotes: 96
Citations: 25

Read summary Original

34

arXiv.org

WebSailor-V2: Bridging the Chasm to Proprietary Agents via Synthetic Data and Scalable Reinforcement Learning

WebSailor-V2 is a post-training pipeline for open-source web agents, comprising data construction (SailorFog-QA-V2), SFT, and RL. The dataset uses a dense knowledge graph with cyclic structures and diverse uncertainty types beyond obfuscation. Training employs a dual-environment RL framework: a high-fidelity simulator for rapid iteration and a managed…

Kuan Li, Zhongwang Zhang, Huifeng Yin, Rui Ye, et al.

Published: Sep 2025
Upvotes: 93
Citations: 60

Read summary Original

35

arXiv.org

Tree Search for LLM Agent Reinforcement Learning

The paper introduces Tree-GRPO, a reinforcement learning method for LLM agents that uses tree search for rollouts, with each node representing a complete agent interaction step (Thought-Action-Observation). This approach shares common prefixes across rollouts, increasing the number of samples under a fixed token/tool-call budget. Tree-GRPO estimates group…

Yuxiang Ji, Ziyu Ma, Yong Wang, Guanhua Chen, et al.

Published: Sep 2025
Upvotes: 92
Citations: 52

Read summary Original

36

arXiv.org

Hala Technical Report: Building Arabic-Centric Instruction & Translation Models at Scale

The HALA technical report introduces a family of Arabic-centric instruction and translation models built with a translate-and-tune pipeline. The pipeline compresses a strong AR↔EN teacher model (CohereLabs/command-a-translate-08-2025) to FP8, achieving ~2x higher throughput with no quality loss. This teacher translates 405K Open-Orca pairs into Arabic, and…

Hasan Abed Al Kader Hammoud, Mohammad Zbeeb, Bernard Ghanem

Published: Sep 2025
Upvotes: 90
Citations: 5

Read summary Original

37

arXiv.org

Seedream 4.0: Toward Next-generation Multimodal Image Generation

Seedream 4.0 is a multimodal image generation system by ByteDance Seed that unifies text-to-image (T2I) synthesis, image editing, and multi-image composition in a single framework. It uses an efficient diffusion transformer (DiT) with a high-compression VAE, reducing image tokens and enabling native 1K-4K resolution generation. The model is pretrained on…

Team Seedream, :, Yunpeng Chen, Yu Gao, et al.

Published: Sep 2025
Upvotes: 88
Citations: 226

Read summary Original

38

arXiv.org

SimpleTIR: End-to-End Reinforcement Learning for Multi-Turn Tool-Integrated Reasoning

SimpleTIR is a plug-and-play algorithm that stabilizes multi-turn Tool-Integrated Reasoning (TIR) training under the Zero RL paradigm. The authors identify that training instability and gradient explosions stem from distributional drift caused by external tool feedback, which leads to the generation of low-probability tokens that compound over turns. They…

Zhenghai Xue, Longtao Zheng, Qian Liu, Yingru Li, et al.

Published: Sep 2025
Upvotes: 84
Citations: 139

Read summary Original

39

arXiv.org

Visual Representation Alignment for Multimodal Large Language Models

Multimodal large language models (MLLMs) trained with visual instruction tuning underperform on vision-centric tasks like object counting and spatial reasoning. The authors attribute this to text-only supervision, which provides indirect guidance for the visual pathway, causing MLLMs to discard fine-grained visual details from the vision encoder. They…

Heeji Yoon, Jaewoo Jung, Junwan Kim, Hyungyu Choi, et al.

Published: Sep 2025
Upvotes: 84
Citations: 35

Read summary Original

40

arXiv.org

WebExplorer: Explore and Evolve for Training Long-Horizon Web Agents

WebExplorer introduces a data synthesis framework for training long-horizon web agents. It uses model-based exploration, where LLMs iteratively search and browse from a seed entity to construct an information space, and iterative long-to-short query evolution, which removes salient clues and adds obfuscation to increase query difficulty. This produces the…

Junteng Liu, Yunji Li, Chi Zhang, Jingyang Li, et al.

Published: Sep 2025
Upvotes: 83
Citations: 74

Read summary Original

41

arXiv.org

VerlTool: Towards Holistic Agentic Reinforcement Learning with Tool Use

VERLTOOL is a unified, modular framework for Agentic Reinforcement Learning with Tool use (ARLT), addressing fragmentation, synchronous execution bottlenecks, and limited extensibility in existing ARLT codebases. It builds on VeRL, providing upstream alignment, a standardized tool server API supporting diverse tools (code execution, search, SQL, vision,…

Dongfu Jiang, Yi Lu, Zhuofeng Li, Zhiheng Lyu, et al.

Published: Sep 2025
Upvotes: 82
Citations: 81

Read summary Original

42

arXiv.org

SimpleVLA-RL: Scaling VLA Training via Reinforcement Learning

SimpleVLA-RL is an efficient reinforcement learning (RL) framework for Vision-Language-Action (VLA) models, built on veRL, that addresses data scarcity and poor generalization in robotic manipulation. It uses interactive trajectory sampling, parallel multi-environment rendering, and outcome-based rewards (1 for success, 0 for failure) with GRPO, enhanced…

Haozhan Li, Yuxin Zuo, Jiale Yu, Yuhao Zhang, et al.

Published: Sep 2025
Upvotes: 82
Citations: 130

Read summary Original

43

arXiv.org

Fathom-DeepResearch: Unlocking Long Horizon Information Retrieval and Synthesis for SLMs

Fathom-DeepResearch is an agentic system for long-horizon information retrieval and synthesis, composed of two 4B-parameter models built on Qwen3-4B. Fathom-Search-4B handles evidence-based investigation via live web search and targeted page querying, trained with DUETQA (a ~5K-sample dataset generated via multi-agent self-play ensuring strict web-search…

Shreyas Singh, Kunal Singh, Pradeep Moturi

Published: Sep 2025
Upvotes: 81
Citations: 4

Read summary Original

44

arXiv.org

More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models

This paper investigates the dual nature of reasoning in Vision-Language Models (VLMs). The authors find that while reasoning improves logical inference, longer reasoning lengths can impair perceptual grounding, leading to errors on basic visual questions. They attribute this to 'visual forgetting,' where models increasingly ignore visual input during…

Xinyu Tian, Shu Zou, Zhaoyuan Yang, Mengqi He, et al.

Published: Sep 2025
Upvotes: 81
Citations: 26

Read summary Original

45

arXiv.org

ReSum: Unlocking Long-Horizon Search Intelligence via Context Summarization

ReSum is a plug-and-play paradigm that enables LLM-based web agents to perform unbounded exploration by periodically invoking an external summarization tool to condense interaction histories into compact summaries, addressing the conflict between extensive exploration and limited context windows. The authors developed ReSumTool-30B, a specialized summary…

Xixi Wu, Kuan Li, Yida Zhao, Liwen Zhang, et al.

Published: Sep 2025
Upvotes: 80
Citations: 95

Read summary Original

46

arXiv.org

SINQ: Sinkhorn-Normalized Quantization for Calibration-Free Low-Precision LLM Weights

SINQ is a calibration-free post-training quantization method for large language models that introduces a dual-scale parameterization (row and column scale vectors) for weight matrices. It uses a Sinkhorn-Knopp-style algorithm to iteratively normalize row and column standard deviations, balancing activation-aware column scaling with row-wise kurtosis…

Lorenz K. Müller, Philippe Bich, Jiawei Zhuang, Ahmet Çelik, et al.

Published: Sep 2025
Upvotes: 80
Citations: 5

Read summary Original

47

arXiv.org

Towards a Unified View of Large Language Model Post-Training

This paper introduces a unified theoretical framework for large language model (LLM) post-training, showing that Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) are not contradictory but instances of a single optimization process. The authors derive a Unified Policy Gradient Estimator (UPGE) that subsumes the gradients of various post-training…

Xingtai Lv, Yuxin Zuo, Youbang Sun, Hongyi Liu, et al.

Published: Sep 2025
Upvotes: 77
Citations: 39

Read summary Original

48

arXiv.org

MemMamba: Rethinking Memory Patterns in State Space Model

MemMamba is a novel architecture for long-sequence modeling that addresses the memory decay problem in state space models (SSMs) like Mamba. The paper first analyzes Mamba's memory mechanism, showing that its state update leads to exponential decay of early information both within layers (horizontal) and across layers (vertical), quantified by new metrics:…

Youjin Wang, Yangjingyi Chen, Jiahao Yan, Jiaxuan Lu, et al.

Published: Sep 2025
Upvotes: 74
Citations: 2

Read summary Original

49

arXiv.org

RewardDance: Reward Scaling in Visual Generation

RewardDance is a scalable reward modeling framework for visual generation that addresses limitations in existing reward models (RMs) by using a generative paradigm. It reformulates reward scoring as the probability of a 'yes' token in a Vision-Language Model (VLM), aligning with the model's next-token prediction mechanism. This enables scaling along two…

Jie Wu, Yu Gao, Zilyu Ye, Ming Li, et al.

Published: Sep 2025
Upvotes: 73
Citations: 58

Read summary Original

50

Annual Meeting of the Association for Computational Linguistics

Towards General Agentic Intelligence via Environment Scaling

This paper introduces AgentScaler, a family of models trained to advance general agentic intelligence through systematic environment scaling. The authors propose a two-stage pipeline: first, they automatically construct diverse, fully simulated environments by collecting over 30,000 APIs, organizing them into domains via community detection, and…

Runnan Fang, Shihao Cai, Baixuan Li, Jialong Wu, et al.

Published: Sep 2025
Upvotes: 73
Citations: 48

Read summary Original

51

arXiv.org

Variational Reasoning for Language Models

The paper introduces a variational reasoning framework for language models, treating thinking traces as latent variables and optimizing them via variational inference. It extends the evidence lower bound (ELBO) to a multi-trace IWAE-style objective and proposes a forward-KL formulation for stable training of the variational posterior. The authors show that…

Xiangxin Zhou, Zichen Liu, Haonan Wang, Chao Du, et al.

Published: Sep 2025
Upvotes: 70
Citations: 6

Read summary Original

52

arXiv.org

Language Models Can Learn from Verbal Feedback Without Scalar Rewards

The paper proposes Feedback-Conditional Policy (FCP), a method for training LLMs directly from verbal feedback without converting it into scalar rewards. FCP treats feedback as a conditioning signal, learning a policy πθ(o|x,c) that approximates the posterior P_off(o|x,c) ∝ π_ref(o|x)·p_env(c|x,o) via maximum likelihood on offline data. An online…

Renjie Luo, Zichen Liu, Xiangyan Liu, Chao Du, et al.

Published: Sep 2025
Upvotes: 70
Citations: 15

Read summary Original

53

arXiv.org

WebResearcher: Unleashing unbounded reasoning capability in Long-Horizon Agents

WebResearcher is a framework for deep-research agents that addresses the limitations of mono-contextual approaches, which suffer from context suffocation and noise contamination. It introduces IterResearch, an iterative paradigm that reformulates deep research as a Markov Decision Process, where agents periodically consolidate findings into evolving…

Zile Qiao, Guoxin Chen, Xuanzhong Chen, Donglei Yu, et al.

Published: Sep 2025
Upvotes: 68
Citations: 66

Read summary Original

54

arXiv.org

Winning the Pruning Gamble: A Unified Approach to Joint Sample and Token Pruning for Efficient Supervised Fine-Tuning

The paper introduces Q-Tuning, a unified framework for joint sample- and token-level pruning during supervised fine-tuning (SFT) of large language models. It addresses the inefficiency of existing methods that treat sample and token pruning separately. The authors propose the Error–Uncertainty (EU) Plane, which categorizes training samples into four…

Shaobo Wang, Jiaming Wang, Jiajun Zhang, Cong Wang, et al.

Published: Sep 2025
Upvotes: 68
Citations: 9

Read summary Original

55

Annual Meeting of the Association for Computational Linguistics

Reinforcement Learning on Pre-Training Data

The paper introduces RLPT (Reinforcement Learning on Pre-Training data), a new training-time scaling paradigm for LLMs that applies reinforcement learning directly to pre-training data, eliminating the need for human annotations. RLPT uses a next-segment reasoning objective where the model predicts a subsequent text segment, and a generative reward model…

Siheng Li, Kejiao Li, Zenan Xu, Guanhua Huang, et al.

Published: Sep 2025
Upvotes: 67
Citations: 11

Read summary Original

56

arXiv.org

OmniInsert: Mask-Free Video Insertion of Any Reference via Diffusion Transformer Models

OmniInsert is a unified framework for mask-free video insertion, allowing users to insert single or multiple reference subjects into a source video based on a text prompt. It addresses three key challenges: data scarcity, subject-scene equilibrium, and insertion harmonization. To tackle data scarcity, the authors propose InsertPipe, a data pipeline with…

Jinshu Chen, Xinghui Li, Xu Bai, Tianxiang Ma, et al.

Published: Sep 2025
Upvotes: 66
Citations: 9

Read summary Original

57

arXiv.org

StableToken: A Noise-Robust Semantic Speech Tokenizer for Resilient SpeechLLMs

StableToken is a semantic speech tokenizer designed to be robust to acoustic noise, addressing the fragility of existing VQ-based tokenizers that produce unstable token sequences under perturbations. The paper identifies two flaws: a brittle single-path quantization architecture and a distant ASR training signal that ignores intermediate token stability.…

Yuhan Song, Linhao Zhang, Chuhan Wu, Aiwei Liu, et al.

Published: Sep 2025
Upvotes: 66
Citations: 4

Read summary Original

58

arXiv.org

MachineLearningLM: Scaling Many-shot In-context Learning via Continued Pretraining

The paper introduces MACHINELEARNINGLM, a continued-pretraining framework that enables large language models (LLMs) to perform many-shot in-context learning (ICL) on tabular machine-learning tasks without gradient updates. The method synthesizes millions of tasks from structural causal models (SCMs), uses a random-forest teacher for a warm-start…

Haoyu Dong, Pengkun Zhang, Mingzhe Lu, Yanzhen Shen, et al.

Published: Sep 2025
Upvotes: 65
Citations: 4

Read summary Original

59

arXiv.org

ReviewScore: Misinformed Peer Review Detection with Large Language Models

The paper introduces REVIEWSCORE, a metric to detect misinformed review points in peer reviews, defined as questions answerable by the paper or weaknesses with incorrect premises. Human annotation of ICLR reviews found 15.2% of weaknesses and 26.4% of questions misinformed. The authors propose an automatic argument reconstruction engine that extracts…

Hyun Ryu, Doohyuk Jang, Hyemin S. Lee, Joonhyun Jeong, et al.

Published: Sep 2025
Upvotes: 64
Citations: 2

Read summary Original

60

arXiv.org

MiniCPM-V 4.5: Cooking Efficient MLLMs via Architecture, Data, and Training Recipe

MiniCPM-V 4.5 is an 8B-parameter multimodal large language model (MLLM) designed for high efficiency and strong performance. It introduces three core improvements: a unified 3D-Resampler architecture that compresses images and videos into compact token sequences (up to 16x for images and an additional 6x for videos), a unified learning paradigm for…

Tianyu Yu, Zefan Wang, Chongyi Wang, Fuwei Huang, et al.

Published: Sep 2025
Upvotes: 63
Citations: 124

Read summary Original

61

arXiv.org

Reasoning Vectors: Transferring Chain-of-Thought Capabilities via Task Arithmetic

This paper introduces the concept of a reasoning vector, which captures the reasoning capability instilled by reinforcement learning (RL) in large language models (LLMs) and can be transferred to other models via simple tensor arithmetic. The authors source two publicly available QWEN2.5 models (1.5B and 7B) with identical initialization, one fine-tuned…

Mohammad Zbeeb, Hasan Abed Al Kader Hammoud, Bernard Ghanem

Published: Sep 2025
Upvotes: 62
Citations: 5

Read summary Original

62

arXiv.org

EchoX: Towards Mitigating Acoustic-Semantic Gap via Echo Training for Speech-to-Speech LLMs

EchoX is a framework for speech-to-speech large language models (SLLMs) that addresses the acoustic-semantic gap, which causes knowledge and reasoning degradation compared to text-based LLMs. The authors propose a three-stage training process: (1) speech-to-text training, (2) text-to-codec training, and (3) Echo training, where a frozen text-to-codec…

Yuhao Zhang, Yuhao Du, Zhanchen Dai, Xiangnan Ma, et al.

Published: Sep 2025
Upvotes: 62
Citations: 3

Read summary Original

63

arXiv.org

Multiplayer Nash Preference Optimization

This paper introduces Multiplayer Nash Preference Optimization (MNPO), a framework that generalizes Nash learning from human feedback (NLHF) to n-player games for aligning large language models with human preferences. MNPO formulates alignment as a game where each policy competes against a population of opponents while being regularized toward a reference…

Fang Wu, Xu Huang, Weihao Xuan, Zhiwei Zhang, et al.

Published: Sep 2025
Upvotes: 62
Citations: 6

Read summary Original

64

arXiv.org

Mini-o3: Scaling Up Reasoning Patterns and Interaction Turns for Visual Search

Mini-o3 is a system for visual search that scales up tool-based interactions and reasoning patterns, achieving state-of-the-art performance on challenging tasks. It addresses limitations of existing open-source VLMs, which show monotonous reasoning and limited interaction turns. The approach includes three key components: constructing the Visual Probe…

Xin Lai, Junyi Li, Wei Li, Tao Liu, et al.

Published: Sep 2025
Upvotes: 60
Citations: 101

Read summary Original

65

NVIDIA

Fast-dLLM v2: Efficient Block-Diffusion LLM

Fast-dLLM v2 is a block diffusion language model that adapts pretrained autoregressive LLMs (Qwen2.5-Instruct 1.5B and 7B) for parallel text generation. It requires only ~1B tokens of fine-tuning, a 500x reduction compared to full-attention diffusion models like Dream (580B tokens). The method uses block-wise diffusion with complementary masking and a…

Chengyue Wu, Hao Zhang, Shuchen Xue, Shizhe Diao, et al.

Published: Sep 2025
Upvotes: 60
Citations: 114

Read summary Original

66

arXiv.org

3D and 4D World Modeling: A Survey

This survey provides the first comprehensive review of 3D and 4D world modeling, addressing the lack of standardized definitions and the fragmented literature that often focuses on 2D generative methods. It establishes precise definitions and a hierarchical taxonomy categorizing methods into video-based (VideoGen), occupancy-based (OccGen), and LiDAR-based…

Lingdong Kong, Yu Yang, Jianbiao Mei, Youquan Liu, et al.

Published: Sep 2025
Upvotes: 59
Citations: 65

Read summary Original

67

AAAI Conference on Artificial Intelligence

DeepResearch Arena: The First Exam of LLMs' Research Abilities via Seminar-Grounded Tasks

The paper introduces DeepResearch Arena, a benchmark for evaluating deep research agents using tasks derived from academic seminar transcripts. It addresses limitations of existing benchmarks, which either risk data leakage from static corpora or lack scalability and realism in expert-curated sets. The authors propose a Multi-Agent Hierarchical Task…

Haiyuan Wan, Chen Yang, Junchi Yu, Meiqi Tu, et al.

Published: Sep 2025
Upvotes: 58
Citations: 26

Read summary Original

68

arXiv.org

Inverse IFEval: Can LLMs Unlearn Stubborn Training Conventions to Follow Real Instructions?

The paper introduces Inverse IFEval, a benchmark to evaluate LLMs' ability to follow counterintuitive instructions that conflict with training conventions. It identifies 'cognitive inertia' from SFT, where models struggle with instructions deviating from standard patterns. The benchmark includes eight instruction types (e.g., Question Correction,…

Qinyan Zhang, Xinping Lei, Ruijie Miao, Yu Fu, et al.

Published: Sep 2025
Upvotes: 58
Citations: 6

Read summary Original

69

arXiv.org

MANZANO: A Simple and Scalable Unified Multimodal Model with a Hybrid Vision Tokenizer

Manzano is a unified multimodal LLM that integrates image understanding and generation using a hybrid vision tokenizer. A single shared vision encoder feeds two lightweight adapters: a continuous adapter for understanding tasks and a discrete adapter (using FSQ with a 64K codebook) for generation, producing tokens in a common semantic space. A unified…

Yanghao Li, Rui Qian, Bowen Pan, Haotian Zhang, et al.

Published: Sep 2025
Upvotes: 58
Citations: 15

Read summary Original

70

arXiv.org

Revolutionizing Reinforcement Learning Framework for Diffusion Large Language Models

The paper introduces TraceRL, a trajectory-aware reinforcement learning framework for diffusion language models (DLMs) that aligns training objectives with inference trajectories, improving reasoning performance on math and coding tasks. It includes a diffusion-based value model for training stability and supports both full-attention and block-attention…

Yinjie Wang, Ling Yang, Bowen Li, Ye Tian, et al.

Published: Sep 2025
Upvotes: 57
Citations: 75

Read summary Original

71

Research paper

Verifying Meta-Awareness via Predictive Rewards in Reasoning Models

The paper introduces MAPR (Meta-Awareness via Predictive Reward), a reinforcement learning framework that enhances reasoning models by training them to predict their own rollout statistics—specifically solution length, pass-rate, and concepts used—and rewarding alignment with actual statistics. This self-generated meta-prediction is verified against…

Yoonjeon Kim, Doohyuk Jang, Eunho Yang

Published: Sep 2025
Upvotes: 57
Citations: 2

Read summary Original

72

arXiv.org

AgentGym-RL: Training LLM Agents for Long-Horizon Decision Making through Multi-Turn Reinforcement Learning

AgentGym-RL is a unified, modular reinforcement learning framework for training LLM agents in multi-turn, long-horizon decision-making tasks, without requiring supervised fine-tuning. It supports diverse environments (web navigation, deep search, digital games, embodied tasks, scientific tasks) and mainstream RL algorithms (PPO, GRPO, REINFORCE++, RLOO).…

Zhiheng Xi, Jixuan Huang, Chenyang Liao, Baodai Huang, et al.

Published: Sep 2025
Upvotes: 56
Citations: 62

Read summary Original

73

Meta AI

TruthRL: Incentivizing Truthful LLMs via Reinforcement Learning

TruthRL is a reinforcement learning framework that directly optimizes the truthfulness of large language models (LLMs) by using a ternary reward scheme that distinguishes correct answers, hallucinations, and abstentions. Implemented with GRPO, it rewards correct answers, penalizes hallucinations, and treats abstentions neutrally, encouraging models to…

Zhepei Wei, Xiao Yang, Kai Sun, Jiaqi Wang, et al.

Published: Sep 2025
Upvotes: 56
Citations: 16

Read summary Original

74

arXiv.org

Set Block Decoding is a Language Model Inference Accelerator

Set Block Decoding (SBD) is a new inference acceleration paradigm for autoregressive language models that integrates next token prediction (NTP) with masked token prediction (MATP) in a single architecture. SBD allows sampling multiple, non-consecutive future tokens in parallel, using solvers from discrete diffusion literature like the Entropy Bounded (EB)…

Itai Gat, Heli Ben-Hamu, Marton Havasi, Daniel Haziza, et al.

Published: Sep 2025
Upvotes: 54
Citations: 18

Read summary Original

75

arXiv.org

No Prompt Left Behind: Exploiting Zero-Variance Prompts in LLM Reinforcement Learning via Entropy-Guided Advantage Shaping

The paper introduces RL-ZVP, a reinforcement learning algorithm for LLMs that exploits zero-variance prompts—where all sampled responses to a prompt receive identical rewards—which standard methods like GRPO discard due to zero advantage. RL-ZVP assigns advantages to these prompts based on response correctness (positive for all-correct, negative for…

Thanh-Long V. Le, Myeongho Jeon, Kim Vu, Viet Lai, et al.

Published: Sep 2025
Upvotes: 54
Citations: 41

Read summary Original

76

NVIDIA

SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer

SANA-Video is a small diffusion model for efficient, high-resolution (up to 720×1280) and minute-long video generation, deployable on RTX 5090 GPUs. It uses a Linear DiT with linear attention (O(N) complexity) and a constant-memory KV cache for block linear attention, enabling long videos with fixed memory. Training cost is 12 days on 64 H100 GPUs (1% of…

Junsong Chen, Yuyang Zhao, Jincheng Yu, Ruihang Chu, et al.

Published: Sep 2025
Upvotes: 54
Citations: 82

Read summary Original

77

arXiv.org

OpenGPT-4o-Image: A Comprehensive Dataset for Advanced Image Generation and Editing

OpenGPT-4o-Image is a large-scale dataset for training unified multimodal models in image generation and editing. It contains 80,000 instruction-image pairs across 11 major domains and 51 subtasks, built using a hierarchical taxonomy and an automated pipeline that leverages GPT-4o. The taxonomy covers five generation modules (Style Control, Complex…

Zhihong Chen, Xuehai Bai, Yang Shi, Chaoyou Fu, et al.

Published: Sep 2025
Upvotes: 54
Citations: 24

Read summary Original

78

arXiv.org

Robix: A Unified Model for Robot Interaction, Reasoning and Planning

Robix is a unified vision-language model that integrates robot reasoning, task planning, and natural language interaction within a single architecture, serving as the high-level cognitive layer in a hierarchical robot system. It generates atomic commands for a low-level controller and verbal responses for humans, enabling complex instruction following,…

Huang Fang, Mengxi Zhang, Heng Dong, Wei Li, et al.

Published: Sep 2025
Upvotes: 53
Citations: 21

Read summary Original

79

arXiv.org

Reasoning over Boundaries: Enhancing Specification Alignment via Test-time Deliberation

The paper introduces specification alignment, a challenge for LLMs to follow dynamic, scenario-specific behavioral and safety specifications. To evaluate this, the authors present SPECBENCH, a benchmark covering 5 scenarios, 103 specifications, and 1,500 prompts. Experiments on 33 models reveal significant alignment gaps and a safety-behavior trade-off.…

Haoran Zhang, Yafu Li, Xuyang Hu, Dongrui Liu, et al.

Published: Sep 2025
Upvotes: 53
Citations: 3

Read summary Original