The Year in AI Papers
The year/March 2026

Monthly collection

March 2026

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

Papers
100
Citations
1.1K
Research labs
9

Explore the year

View all months →
Aug2025, 63 papersSep2025, 79 papersOct2025, 98 papersNov2025, 73 papersDec2025, 66 papersJan2026, 86 papersFeb2026, 86 papersMar2026, current month, 100 papersApr2026, 77 papersMay2026, 115 papersJun2026, 86 papersJul2026, 84 papersAug2026, 3 papers
1,016papers
curated and indexed

100 papers published in March 2026

01

arXiv.org

AI Can Learn Scientific Taste

The paper introduces Reinforcement Learning from Community Feedback (RLCF), a paradigm that uses citation-based community signals to train AI models for scientific taste—the ability to judge and propose high-impact research ideas. They built SciJudgeBench with 720K field- and time-matched paper abstract pairs, trained Scientific Judge via GRPO to predict…

Jingqi Tong, Mingzhe Li, Hangcheng Li, Yongzhuo Yang, et al.
Published
Mar 2026
Upvotes
432
Citations
4
Read summary Original
02

Research paper

Demystifying Video Reasoning

This paper challenges the Chain-of-Frames (CoF) hypothesis for reasoning in diffusion-based video models, proposing instead that reasoning primarily emerges along the diffusion denoising steps, termed Chain-of-Steps (CoS). Through qualitative analysis and noise perturbation experiments, the authors show that models explore multiple candidate solutions in…

Ruisi Wang, Zhongang Cai, Fanyi Pu, Junxiang Xu, et al.
Published
Mar 2026
Upvotes
373
Citations
6
Read summary Original
03

arXiv.org

DataFlex: A Unified Framework for Data-Centric Dynamic Training of Large Language Models

DataFlex is a unified data-centric dynamic training framework built on LLaMA-Factory, designed to treat data as a first-class optimization variable. It integrates three paradigms—data selection, data mixture optimization, and data reweighting—through modular trainer abstractions (Select, Mix, Weight Trainers) and pluggable components (selectors, mixers,…

Hao Liang, Zhengyang Zhao, Meiyi Qiang, Mingrui Chen, et al.
Published
Mar 2026
Upvotes
365
Citations
3
Read summary Original
04

Qwen

FIPO: Eliciting Deep Reasoning with Future-KL Influenced Policy Optimization

FIPO (Future-KL Influenced Policy Optimization) is a reinforcement learning algorithm that addresses the coarse-grained credit assignment problem in GRPO-style training for large language models. By incorporating a discounted Future-KL divergence into the policy update, FIPO re-weights token advantages based on their influence on subsequent trajectory…

Chiyu Ma, Shuo Yang, Kexin Huang, Jinda Lu, et al.
Published
Mar 2026
Upvotes
353
Citations
18
Read summary Original
05

arXiv.org

CARLA-Air: Fly Drones Inside a CARLA World -- A Unified Infrastructure for Air-Ground Embodied Intelligence

CARLA-Air is an open-source simulation platform that integrates CARLA and AirSim within a single Unreal Engine process, enabling joint simulation of aerial and ground agents for embodied intelligence research. It resolves the UE4 single-game-mode conflict by inheriting CARLA's ground subsystems and composing AirSim's flight actor as a world entity,…

Tianle Zeng, Yanci Wen, Hong Zhang
Published
Mar 2026
Upvotes
344
Citations
2
Read summary Original
06

arXiv.org

InCoder-32B: Code Foundation Model for Industrial Scenarios

InCoder-32B is a 32B-parameter code foundation model designed to unify code intelligence across general programming and industrial domains such as chip design, GPU kernel optimization, embedded systems, compiler optimization, and 3D modeling. It is trained from scratch using a three-stage Code-Flow pipeline: pre-training with curated industrial data,…

Jian Yang, Wei Zhang, Jiajun Wu, Junhang Cheng, et al.
Published
Mar 2026
Upvotes
312
Citations
6
Read summary Original
07

arXiv.org

SocialOmni: Benchmarking Audio-Visual Social Interactivity in Omni Models

SocialOmni is a new benchmark for evaluating social interactivity in omni-modal large language models (OLMs), addressing the gap left by static, accuracy-centric benchmarks. It evaluates three dimensions: who is speaking (speaker identification), when to interject (turn-taking timing), and how to phrase interruptions (response generation). The benchmark…

Tianyu Xie, Jinfa Huang, Yuexiao Ma, Rongfang Luo, et al.
Published
Mar 2026
Upvotes
248
Citations
5
Read summary Original
08

arXiv.org

Bootstrapping Exploration with Group-Level Natural Language Feedback in Reinforcement Learning

The paper introduces GOLF, a reinforcement learning (RL) framework that leverages group-level natural language (NL) feedback to improve exploration efficiency in training large language models (LLMs). Unlike standard RL that relies solely on scalar rewards, GOLF aggregates two complementary feedback sources: external critiques (pinpointing errors or…

Lei Huang, Xiang Cheng, Chenxiao Zhao, Guobin Shen, et al.
Published
Mar 2026
Upvotes
211
Citations
2
Read summary Original
09

arXiv.org

Heterogeneous Agent Collaborative Reinforcement Learning

The paper introduces Heterogeneous Agent Collaborative Reinforcement Learning (HACRL), a new RLVR problem where heterogeneous LLM agents share verified rollouts during training to mutually improve, while executing independently at inference. Unlike LLM-based MARL, HACRL does not require coordinated deployment, and unlike distillation, it enables…

Zhixia Zhang, Zixuan Huang, Gongxun Li, Huaiyang Wang, et al.
Published
Mar 2026
Upvotes
199
Citations
12
Read summary Original
10

Moonshot AI

Attention Residuals

The paper proposes Attention Residuals (AttnRes), replacing fixed unit-weight residual connections in LLMs with learned, input-dependent softmax attention over preceding layer outputs. This addresses PreNorm dilution, where hidden-state magnitudes grow with depth, diluting layer contributions. A scalable variant, Block AttnRes, partitions layers into…

Kimi Team, Guangyu Chen, Yu Zhang, Jianlin Su, et al.
Published
Mar 2026
Upvotes
194
Citations
42
Read summary Original
11

arXiv.org

Helios: Real Real-Time Long Video Generation Model

Helios is a 14B autoregressive diffusion model for video generation that achieves 19.5 FPS on a single NVIDIA H100 GPU, enabling minute-scale generation without common anti-drifting heuristics (e.g., self-forcing, error-banks) or standard acceleration techniques (e.g., KV-cache, quantization). It unifies T2V, I2V, and V2V tasks via a unified input…

Shenghai Yuan, Yuanyang Yin, Zongjian Li, Xinwei Huang, et al.
Published
Mar 2026
Upvotes
190
Citations
33
Read summary Original
12

arXiv.org

Utonia: Toward One Encoder for All Point Clouds

Utonia introduces a single self-supervised point transformer encoder trained jointly on diverse point cloud domains, including indoor scans, outdoor LiDAR, remote sensing, object CAD, and video-lifted point clouds. The authors identify three core cross-domain mismatches: inconsistent modality availability, sensitivity to granularity shifts, and bias toward…

Yujia Zhang, Xiaoyang Wu, Yunhan Yang, Xianzhe Fan, et al.
Published
Mar 2026
Upvotes
187
Citations
9
Read summary Original
13

arXiv.org

MiroThinker-1.7 & H1: Towards Heavy-Duty Research Agents via Verification

MiroMind Team introduces MiroThinker-1.7, a research agent for complex long-horizon reasoning, and MiroThinker-H1, which adds verification-centric reasoning. MiroThinker-1.7 improves step-level reliability through agentic mid-training that emphasizes planning, reasoning, and tool interaction. MiroThinker-H1 integrates local verification (evaluating…

MiroMind Team, S. Bai, L. Bing, L. Lei, et al.
Published
Mar 2026
Upvotes
187
Citations
20
Read summary Original
14

arXiv.org

ClawKeeper: Comprehensive Safety Protection for OpenClaw Agents Through Skills, Plugins, and Watchers

ClawKeeper is a comprehensive security framework for OpenClaw, an open-source autonomous agent runtime, addressing critical vulnerabilities such as sensitive data leakage, privilege escalation, and malicious skill execution. It integrates three complementary protection layers: skill-based protection at the instruction level, plugin-based runtime…

Songyang Liu, Chaozhuo Li, Chenxu Wang, Jinyu Hou, et al.
Published
Mar 2026
Upvotes
183
Citations
18
Read summary Original
15

arXiv.org

OpenClaw-RL: Train Any Agent Simply by Talking

OpenClaw-RL is a framework that enables online reinforcement learning for LLM-based agents by extracting training signals from the next-state signals that follow each agent action (e.g., user replies, tool outputs, GUI changes). It uses a server–client architecture where the RL server hosts the policy behind an inference API and user terminals stream…

Yinjie Wang, Xuyang Chen, Xiaolong Jin, Mengdi Wang, et al.
Published
Mar 2026
Upvotes
158
Citations
66
Read summary Original
16

arXiv.org

Out of Sight but Not Out of Mind: Hybrid Memory for Dynamic Video World Models

The paper introduces Hybrid Memory, a paradigm for video world models that requires maintaining static background consistency while tracking dynamic subjects during out-of-view intervals. The authors construct HM-World, a large-scale dataset of 59K high-fidelity clips with 17 scenes, 49 subjects, and designed exit-entry events, and propose HyDRA, a memory…

Kaijin Chen, Dingkang Liang, Xin Zhou, Yikang Ding, et al.
Published
Mar 2026
Upvotes
157
Citations
11
Read summary Original
17

arXiv.org

Qianfan-OCR: A Unified End-to-End Model for Document Intelligence

Qianfan-OCR is a 4B-parameter end-to-end vision-language model that unifies document parsing, layout analysis, and understanding, outperforming all end-to-end models on OmniDocBench v1.5 (93.12) and OlmOCR Bench (79.8). It introduces Layout-as-Thought, an optional thinking phase triggered by ⟨think⟩ tokens that generates structured layout representations…

Daxiang Dong, Mingming Zheng, Dong Xu, Chunhua Luo, et al.
Published
Mar 2026
Upvotes
155
Citations
11
Read summary Original
18

arXiv.org

ShotStream: Streaming Multi-Shot Video Generation for Interactive Storytelling

ShotStream is a novel causal multi-shot video generation architecture that enables interactive storytelling and real-time synthesis at 16 FPS on a single GPU. It reformulates multi-shot generation as a next-shot prediction task, allowing users to guide narratives via streaming prompts. The method first fine-tunes a text-to-video model into a bidirectional…

Yawen Luo, Xiaoyu Shi, Junhao Zhuang, Yutian Chen, et al.
Published
Mar 2026
Upvotes
155
Citations
14
Read summary Original
19

arXiv.org

Grounding World Simulation Models in a Real-World Metropolis

The paper introduces Seoul World Model (SWM), a city-scale world model that grounds autoregressive video generation in the real city of Seoul via retrieval-augmented conditioning on street-view images. SWM fine-tunes Cosmos-Predict2.5-2B on 440k Seoul street-view images, real driving videos, and synthetic urban data. It addresses three challenges: temporal…

Junyoung Seo, Hyunwook Choi, Minkyung Kwon, Jinhyeok Choi, et al.
Published
Mar 2026
Upvotes
154
Citations
6
Read summary Original
20

arXiv.org

Efficient Reasoning with Balanced Thinking

Large Reasoning Models (LRMs) often overthink (redundant steps on simple problems) or underthink (insufficient exploration), causing inefficiency and inaccuracy. Existing mitigation methods, like suppressing reflection keywords, can induce underthinking. The paper proposes REBALANCE, a training-free framework to balance these extremes. It uses stepwise…

Yulin Li, Tengyao Tu, Li Ding, Junjie Wang, et al.
Published
Mar 2026
Upvotes
152
Citations
3
Read summary Original
21

arXiv.org

HSImul3R: Physics-in-the-Loop Reconstruction of Simulation-Ready Human-Scene Interactions

HSImul3R is a unified framework for reconstructing simulation-ready 3D human–scene interactions (HSI) from casual captures like sparse-view images and monocular videos. It addresses the perception–simulation gap where visually plausible reconstructions fail in physics engines. The method uses a physically-grounded bi-directional optimization pipeline with…

Yukang Cao, Haozhe Xie, Fangzhou Hong, Long Zhuo, et al.
Published
Mar 2026
Upvotes
152
Citations
2
Read summary Original
22

arXiv.org

OmniLottie: Generating Vector Animations via Parameterized Lottie Tokens

OmniLottie is a versatile framework for generating high-quality vector animations from multi-modal instructions (text, image, and video). It addresses the challenge of raw Lottie JSON files containing extensive invariant structural metadata by introducing a Lottie tokenizer that converts JSON into compact sequences of commands and parameters. This…

Yiying Yang, Wei Cheng, Sijin Chen, Honghao Fu, et al.
Published
Mar 2026
Upvotes
151
Citations
0
Read summary Original
23

arXiv.org

EnterpriseOps-Gym: Environments and Evaluations for Stateful Agentic Planning and Tool Use in Enterprise Settings

ENTERPRISEOPS-GYM is a benchmark for evaluating LLM agents in realistic enterprise settings, featuring a containerized sandbox with 164 database tables and 512 tools across eight domains (CSM, HR, ITSM, Email, Calendar, Teams, Drive, Hybrid). It includes 1,150 expert-curated tasks, 30 of which are infeasible to test refusal behavior, with SQL-based…

Shiva Krishna Reddy Malay, Shravan Nayak, Jishnu Sethumadhavan Nair, Sagar Davasam, et al.
Published
Mar 2026
Upvotes
150
Citations
6
Read summary Original
24

arXiv.org

OpenSeeker: Democratizing Frontier Search Agents by Fully Open-Sourcing Training Data

OpenSeeker is the first fully open-source search agent (model and data) achieving frontier-level performance, developed by an academic team at Shanghai Jiao Tong University. It uses two core innovations: fact-grounded scalable controllable QA synthesis, which reverse-engineers the web graph via topological expansion and entity obfuscation to generate…

Yuwen Du, Rui Ye, Shuo Tang, Xinyu Zhu, et al.
Published
Mar 2026
Upvotes
150
Citations
21
Read summary Original
25

arXiv.org

LongCat-Next: Lexicalizing Modalities as Discrete Tokens

LongCat-Next, developed by Meituan's LongCat team, introduces the Discrete Native Autoregressive (DiNA) paradigm, which unifies text, vision, and audio into a shared discrete token space, enabling a single autoregressive model to handle all modalities. A key innovation is the Discrete Native Any-resolution Visual Transformer (dNaViT), which uses…

Meituan LongCat Team, Bin Xiao, Chao Wang, Chengjiang Li, et al.
Published
Mar 2026
Upvotes
150
Citations
18
Read summary Original
26

arXiv.org

TAPS: Task Aware Proposal Distributions for Speculative Sampling

This paper investigates how the training distribution of draft models affects speculative decoding performance, using HASS and EAGLE-2 drafters trained on MathInstruct, ShareGPT, and mixed data, evaluated on MT-Bench, GSM8K, MATH-500, and SVAMP. Results show task-specific training yields clear specialization: MathInstruct-trained drafts excel on reasoning…

Mohamad Zbib, Mohamad Bazzi, Ammar Mohanna, Hasan Abed Al Kader Hammoud, et al.
Published
Mar 2026
Upvotes
147
Citations
0
Read summary Original
27

Research paper

Edit in 2D, Verify in 3D: Reinforcement Learning for Multi-view Consistent Scene Editing

The paper introduces RL3DEdit, a single-pass framework for multi-view consistent 3D scene editing that uses reinforcement learning (RL) to augment a 2D diffusion editor (FLUX-Kontext) with 3D consistency priors. The authors observe that while generating multi-view consistent images is challenging, verifying consistency is tractable, making RL a suitable…

Jiyuan Wang, Chunyu Lin, Lei Sun, Zhi Cao, et al.
Published
Mar 2026
Upvotes
145
Citations
7
Read summary Original
28

arXiv.org

MetaClaw: Just Talk -- An Agent That Meta-Learns and Evolves in the Wild

MetaClaw is a continual meta-learning framework that enables deployed LLM agents to evolve through two complementary mechanisms: skill-driven fast adaptation and opportunistic policy optimization. Skill-driven fast adaptation analyzes failure trajectories to synthesize new behavioral instructions, which are immediately injected into the agent's prompt…

Peng Xia, Jianwen Chen, Xinyu Yang, Haoqin Tu, et al.
Published
Mar 2026
Upvotes
141
Citations
20
Read summary Original
29

arXiv.org

MinerU-Diffusion: Rethinking Document OCR as Inverse Rendering via Diffusion Decoding

MinerU-Diffusion is a 2.5B-parameter diffusion-based framework for document OCR that replaces autoregressive (AR) decoding with block-wise parallel diffusion denoising under visual conditioning. The authors argue that left-to-right causal generation is an artifact of serialization, not intrinsic to OCR, and propose inverse rendering via diffusion. The…

Hejun Dong, Junbo Niu, Bin Wang, Weijun Zeng, et al.
Published
Mar 2026
Upvotes
139
Citations
7
Read summary Original
30

arXiv.org

Intern-S1-Pro: Scientific Multimodal Foundation Model at Trillion Scale

Intern-S1-Pro, developed by Shanghai AI Laboratory, is the first one-trillion-parameter scientific multimodal foundation model, built on the SAGE framework. It scales from Intern-S1 via expert expansion with Grouped Routing to ensure load balance and training stability, and uses a Straight-Through Estimator for efficient router updates. The model…

Yicheng Zou, Dongsheng Zhu, Lin Zhu, Tong Zhu, et al.
Published
Mar 2026
Upvotes
134
Citations
14
Read summary Original
31

arXiv.org

Omni-WorldBench: Towards a Comprehensive Interaction-Centric Evaluation for World Models

Omni-WorldBench is a new benchmark for evaluating the interactive response capabilities of video-based world models, addressing the gap left by existing benchmarks that focus on visual fidelity or static 3D reconstruction. It comprises Omni-WorldSuite, a set of 1,068 prompts with initial frames and optional camera trajectories, organized into three…

Meiqi Wu, Zhixin Cai, Fufangchen Zhao, Xiaokun Feng, et al.
Published
Mar 2026
Upvotes
127
Citations
5
Read summary Original
32

arXiv.org

QuanBench+: A Unified Multi-Framework Benchmark for LLM-Based Quantum Code Generation

QuanBench+ is a unified benchmark for evaluating LLMs on quantum code generation across Qiskit, PennyLane, and Cirq, holding task intent constant while varying the framework. It includes 42 tasks spanning quantum algorithms, gate decomposition, and state preparation, and uses executable functional tests with Pass@1, Pass@5, and KL-divergence-based…

Ali Slim, Haydar Hamieh, Jawad Kotaich, Yehya Ghosn, et al.
Published
Mar 2026
Upvotes
126
Citations
1
Read summary Original
33

arXiv.org

Speed by Simplicity: A Single-Stream Architecture for Fast Audio-Video Generative Foundation Model

daVinci-MagiHuman is an open-source audio-video generative foundation model for human-centric generation, jointly producing synchronized video and audio via a single-stream Transformer that processes text, video, and audio in a unified token sequence using self-attention only. This design avoids multi-stream complexity and supports multilingual generation…

SII-GAIR, Sand. ai, :, Ethan Chern, et al.
Published
Mar 2026
Upvotes
125
Citations
12
Read summary Original
34

arXiv.org

T2S-Bench & Structure-of-Thought: Benchmarking and Prompting Comprehensive Text-to-Structure Reasoning

The paper introduces Structure of Thought (SoT), a prompting technique that guides LLMs to construct intermediate text structures (nodes and links) before answering, consistently improving performance across eight text-processing tasks and three model families. Building on this, the authors present T2S-Bench, the first benchmark for evaluating…

Qinsi Wang, Hancheng Ye, Jinhee Kim, Jinghan Ke, et al.
Published
Mar 2026
Upvotes
122
Citations
1
Read summary Original
35

arXiv.org

Penguin-VL: Exploring the Efficiency Limits of VLM with LLM-based Vision Encoders

Penguin-VL, developed by Tencent AI Lab, introduces compact 2B and 8B vision-language models that challenge the reliance on contrastive pretraining (e.g., CLIP/SigLIP) for vision encoders. The authors argue that contrastive learning suppresses fine-grained visual cues needed for reasoning. Instead, they propose Penguin-Encoder, initialized from a text-only…

Boqiang Zhang, Lei Ke, Ruihan Yang, Qi Gao, et al.
Published
Mar 2026
Upvotes
120
Citations
4
Read summary Original
36

arXiv.org

PixelSmile: Toward Fine-Grained Facial Expression Editing

PixelSmile is a diffusion-based framework for fine-grained facial expression editing, addressing semantic overlap between expressions like fear-surprise and anger-disgust. The authors construct the Flex Facial Expression (FFE) dataset with 60,000 images (real and anime) annotated with continuous 12-dimensional affective scores, and establish FFE-Bench to…

Jiabin Hua, Hengyuan Xu, Aojie Li, Wei Cheng, et al.
Published
Mar 2026
Upvotes
118
Citations
0
Read summary Original
37

Qwen

HopChain: Multi-Hop Data Synthesis for Generalizable Vision-Language Reasoning

The paper introduces HopChain, a scalable framework for synthesizing multi-hop vision-language reasoning data for reinforcement learning with verifiable rewards (RLVR) training of vision-language models (VLMs). The authors identify that long chain-of-thought (CoT) reasoning in VLMs suffers from diverse, compounding failure modes (perception, reasoning,…

Shenzhi Wang, Shixuan Liu, Jing Zhou, Chang Gao, et al.
Published
Mar 2026
Upvotes
110
Citations
2
Read summary Original
38

arXiv.org

Astrolabe: Steering Forward-Process Reinforcement Learning for Distilled Autoregressive Video Models

Astrolabe is an online reinforcement learning (RL) framework designed to align distilled autoregressive (AR) video models with human preferences without re-distillation. It addresses limitations of existing RL methods, which require expensive reverse-process optimization or lack active exploration. Astrolabe uses a forward-process RL formulation based on…

Songchun Zhang, Zeyue Xue, Siming Fu, Jie Huang, et al.
Published
Mar 2026
Upvotes
109
Citations
10
Read summary Original
39

Meta AI

Beyond Language Modeling: An Exploration of Multimodal Pretraining

This paper presents controlled, from-scratch experiments to clarify the design space of unified multimodal pretraining, using the Transfusion framework (next-token prediction for language, diffusion for vision) on text, video, image-text pairs, and action-conditioned video. Key findings: (1) Representation Autoencoders (RAE), e.g., SigLIP 2, provide a…

Shengbang Tong, David Fan, John Nguyen, Ellis Brown, et al.
Published
Mar 2026
Upvotes
107
Citations
21
Read summary Original
40

arXiv.org

OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis

OpenResearcher is a fully open pipeline for synthesizing long-horizon deep research trajectories. It decouples one-time corpus bootstrapping from multi-turn trajectory synthesis, running the search-and-browse loop entirely offline over a 15M-document corpus using three browser primitives: search, open, and find. Using GPT-OSS-120B as the teacher, it…

Zhuofeng Li, Dongfu Jiang, Xueguang Ma, Haoxiang Zhang, et al.
Published
Mar 2026
Upvotes
102
Citations
20
Read summary Original
41

arXiv.org

CUA-Suite: Massive Human-annotated Video Demonstrations for Computer-Use Agents

CUA-SUITE is a large-scale ecosystem of human-annotated video demonstrations for training and evaluating desktop computer-use agents (CUAs). It addresses the scarcity of continuous, high-quality human demonstration videos, which recent work identifies as critical for scaling agents. The core resource, VIDEOCUA, provides approximately 10,000…

Xiangru Jian, Shravan Nayak, Kevin Qinghong Lin, Aarash Feizi, et al.
Published
Mar 2026
Upvotes
99
Citations
3
Read summary Original
42

arXiv.org

Terminal Agents Suffice for Enterprise Automation

The paper argues that minimal terminal-based coding agents, which interact directly with platform APIs via a terminal and filesystem, can match or outperform more complex agent architectures for enterprise automation. The authors introduce StarShell, a terminal agent, and compare it against GUI-driven web agents and MCP-based tool-augmented agents across…

Patrice Bechard, Orlando Marquez Ayala, Emily Chen, Jordan Skelton, et al.
Published
Mar 2026
Upvotes
98
Citations
5
Read summary Original
43

arXiv.org

Thinking in Uncertainty: Mitigating Hallucinations in MLRMs with Latent Entropy-Aware Decoding

The paper introduces Latent Entropy-Aware Decoding (LEAD), a plug-and-play decoding strategy to mitigate hallucinations in multimodal large reasoning models (MLRMs). The authors observe that transition words (e.g., because, however, wait) are associated with high-entropy states and hallucinations. They propose entropy-aware reasoning mode switching: in…

Zhongxing Xu, Zhonghua Wang, Zhe Qian, Dachuan Shi, et al.
Published
Mar 2026
Upvotes
95
Citations
10
Read summary Original
44

arXiv.org

Generation Models Know Space: Unleashing Implicit 3D Priors for Scene Understanding

VEGA-3D is a plug-and-play framework that repurposes pre-trained video generation models as Latent World Simulators to provide implicit 3D priors for Multimodal Large Language Models (MLLMs), addressing their spatial blindness. The method extracts spatiotemporal features from intermediate noise levels of a frozen video diffusion model (e.g., Wan2.1-T2V)…

Xianjin Wu, Dingkang Liang, Tianrui Feng, Kui Xia, et al.
Published
Mar 2026
Upvotes
95
Citations
10
Read summary Original
45

Annual Meeting of the Association for Computational Linguistics

Lost in Stories: Consistency Bugs in Long Story Generation by LLMs

The paper introduces ConStory-Bench, a benchmark for evaluating narrative consistency in long-form story generation by LLMs, and CONSTORY-CHECKER, an automated pipeline that detects contradictions with textual evidence. The benchmark includes 2,000 prompts across four task scenarios and a taxonomy of five error categories with 19 subtypes. Evaluation of…

Junjie Li, Xinrui Guo, Yuhao Wu, Roy Ka-Wei Lee, et al.
Published
Mar 2026
Upvotes
93
Citations
3
Read summary Original
46

arXiv.org

WildWorld: A Large-Scale Dataset for Dynamic World Modeling with Actions and Explicit State toward Generative ARPG

WildWorld is a large-scale, action-conditioned world modeling dataset automatically collected from the AAA game Monster Hunter: Wilds. It contains over 108 million frames with more than 450 actions (movement, attacks, skill casting) and per-frame annotations including character skeletons, world states, camera poses, and depth maps. The dataset addresses…

Zhen Li, Zian Meng, Shuwei Shi, Wenshuo Peng, et al.
Published
Mar 2026
Upvotes
92
Citations
6
Read summary Original
47

arXiv.org

Towards a Medical AI Scientist

The paper introduces Medical AI Scientist, the first autonomous research framework tailored to clinical medicine. It generates clinically grounded ideas by transforming surveyed literature into actionable evidence via a clinician-engineer co-reasoning mechanism, and drafts evidence-grounded manuscripts following a structured medical writing paradigm and…

Hongtao Wu, Boyun Zheng, Dingjie Song, Yu Jiang, et al.
Published
Mar 2026
Upvotes
92
Citations
4
Read summary Original
48

arXiv.org

Spatial-TTT: Streaming Visual-based Spatial Intelligence with Test-Time Training

Spatial-TTT is a framework for streaming visual-based spatial intelligence that uses test-time training (TTT) to maintain adaptive fast weights as a compact memory for accumulating 3D evidence from long-horizon videos. It employs a hybrid architecture interleaving TTT layers with self-attention anchor layers at a 3:1 ratio, large-chunk updates, and…

Fangfu Liu, Diankun Wu, Jiawei Chi, Yimo Cai, et al.
Published
Mar 2026
Upvotes
91
Citations
3
Read summary Original
49

arXiv.org

Video-CoE: Reinforcing Video Event Prediction via Chain of Events

The paper introduces Video-CoE, a method to improve video event prediction (VEP) in multimodal large language models (MLLMs). The authors first evaluate leading MLLMs on VEP benchmarks, identifying two main failure causes: lack of logical reasoning for future events and insufficient use of visual information. To address these, they propose the Chain of…

Qile Su, Jing Tang, Rui Chen, Lei Sun, et al.
Published
Mar 2026
Upvotes
91
Citations
2
Read summary Original
50

arXiv.org

Look Where It Matters: High-Resolution Crops Retrieval for Efficient VLMs

The paper presents AwaRes, a framework for efficient vision-language model (VLM) inference that processes a low-resolution global image and uses tool-calling to retrieve only the high-resolution crops needed for a query. AwaRes trains a coupled-decision policy (CDP) that jointly decides whether to escalate resolution and which crops to request. Supervision…

Nimrod Shabtay, Moshe Kimhi, Artem Spector, Sivan Haray, et al.
Published
Mar 2026
Upvotes
91
Citations
1
Read summary Original
51

arXiv.org

MOOSE-Star: Unlocking Tractable Training for Scientific Discovery by Breaking the Complexity Barrier

The paper introduces MOOSE-Star, a framework for tractable training of LLMs for scientific discovery by directly modeling P(hypothesis|background). The authors argue that end-to-end training is intractable due to combinatorial complexity O(N^k) from retrieving k inspirations from a large knowledge base. MOOSE-Star decomposes this into sequential subtasks:…

Zonglin Yang, Lidong Bing
Published
Mar 2026
Upvotes
90
Citations
3
Read summary Original
52

arXiv.org

MosaicMem: Hybrid Spatial Memory for Controllable Video World Models

MosaicMem is a hybrid spatial memory mechanism for video world models that combines explicit 3D structure with implicit, attention-based conditioning. It lifts video patches into 3D for precise localization and retrieval, then composes them in the queried view via a patch-and-compose interface, allowing the model to preserve persistent elements while…

Wei Yu, Runjia Qian, Yumeng Li, Liquan Wang, et al.
Published
Mar 2026
Upvotes
89
Citations
10
Read summary Original
53

arXiv.org

UniG2U-Bench: Do Unified Models Advance Multimodal Understanding?

UniG2U-Bench is a new benchmark for evaluating whether unified multimodal models (UMMs) benefit from generation when performing understanding tasks. It includes 3,000 samples across 7 categories and 30 subtasks, and evaluates over 30 models, including base VLMs, unified models, and agentic models. The study finds that unified models generally underperform…

Zimo Wen, Boxiu Li, Wanbo Zhang, Junxiang Lei, et al.
Published
Mar 2026
Upvotes
88
Citations
3
Read summary Original
54

arXiv.org

GEMS: Agent-Native Multimodal Generation with Memory and Skills

GEMS is a framework for multimodal generation that uses an agent-native approach to improve performance on complex instructions and specialized tasks. It has three main parts: an Agent Loop that iteratively refines generation through planning, decomposition, generation, verification, and refinement; Agent Memory that stores a persistent, hierarchically…

Zefeng He, Siyuan Huang, Xiaoye Qu, Yafu Li, et al.
Published
Mar 2026
Upvotes
88
Citations
8
Read summary Original
55

arXiv.org

Holi-Spatial: Evolving Video Streams into Holistic 3D Spatial Intelligence

Holi-Spatial is a fully automated pipeline that converts raw video streams into holistic 3D spatial annotations without human intervention, addressing the scalability limits of existing spatial intelligence datasets that rely on manually annotated 3D scans. The pipeline has three stages: geometric optimization using 3D Gaussian Splatting (3DGS) with…

Yuanyuan Gao, Hao Li, Yifei Liu, Xinhao Ji, et al.
Published
Mar 2026
Upvotes
87
Citations
5
Read summary Original
56

arXiv.org

Flash-KMeans: Fast and Memory-Efficient Exact K-Means

Flash-KMeans is a GPU implementation of exact k-means that addresses performance bottlenecks in modern AI workloads. The paper identifies two main kernel-level issues: the assignment stage suffers from an IO bottleneck due to materializing the N×K distance matrix in HBM, and the centroid update stage suffers from atomic write contention due to…

Shuo Yang, Haocheng Xi, Yilong Zhao, Muyang Li, et al.
Published
Mar 2026
Upvotes
85
Citations
5
Read summary Original
57

arXiv.org

Mixture-of-Depths Attention

The paper introduces Mixture-of-Depths Attention (MoDA), a mechanism for large language models that allows each attention head to attend to both sequence KV pairs at the current layer and depth KV pairs from preceding layers, addressing information dilution in deep Transformers. MoDA uses a unified softmax over sequence and depth attention, and a…

Lianghui Zhu, Yuxin Fang, Bencheng Liao, Shijie Wang, et al.
Published
Mar 2026
Upvotes
83
Citations
3
Read summary Original
58

arXiv.org

LIBERO-Para: A Diagnostic Benchmark and Metrics for Paraphrase Robustness in VLA Models

The paper introduces LIBERO-Para, a benchmark for evaluating paraphrase robustness in Vision-Language-Action (VLA) models, and PRIDE, a metric for fine-grained robustness analysis. LIBERO-Para varies action expressions and object references independently, yielding 43 paraphrase types. Across seven VLA configurations (0.6B–7.5B), consistent performance…

Chanyoung Kim, Minwoo Kim, Minseok Kang, Hyunwoo Kim, et al.
Published
Mar 2026
Upvotes
83
Citations
3
Read summary Original
59

arXiv.org

Lingshu-Cell: A generative cellular world model for transcriptome modeling toward virtual cells

Lingshu-Cell is a masked discrete diffusion model (MDDM) for generative modeling of single-cell transcriptomics, introduced by Alibaba DAMO Academy. It models transcriptomic state distributions across ~18,000 genes without prior gene selection, operating directly in a discrete token space compatible with sparse, non-sequential scRNA-seq data. The model…

Han Zhang, Guo-Hua Yuan, Chaohao Yuan, Tingyang Xu, et al.
Published
Mar 2026
Upvotes
79
Citations
3
Read summary Original
60

arXiv.org

LongCat-Flash-Prover: Advancing Native Formal Reasoning via Agentic Tool-Integrated Reinforcement Learning

LongCat-Flash-Prover is a 560-billion-parameter open-source Mixture-of-Experts (MoE) model for native formal reasoning in Lean4, developed by Meituan LongCat Team. It decomposes formal reasoning into auto-formalization, sketching, and proving, and uses a Hybrid-Experts Iteration Framework to synthesize high-quality trajectories with tool feedback. The…

Jianing Wang, Jianfei Zhang, Qi Guo, Linsen Guo, et al.
Published
Mar 2026
Upvotes
78
Citations
5
Read summary Original
61

Google DeepMind

Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs

The paper investigates why reasoning improves parametric knowledge recall in LLMs for simple, single-hop factual questions. Using hybrid models (Gemini-2.5-Flash, Gemini-2.5-Pro, Qwen3-32B) on SimpleQA-Verified and EntityQuestions, the authors find that enabling reasoning substantially expands the model's capability boundary, as measured by pass@k, with…

Zorik Gekhman, Roee Aharoni, Eran Ofek, Mor Geva, et al.
Published
Mar 2026
Upvotes
76
Citations
9
Read summary Original
62

arXiv.org

LMEB: Long-horizon Memory Embedding Benchmark

The paper introduces LMEB, a benchmark for evaluating embedding models on long-horizon memory retrieval tasks, which are underexplored in existing benchmarks like MTEB. LMEB comprises 22 datasets and 193 zero-shot retrieval tasks across four memory types: episodic, dialogue, semantic, and procedural. The authors evaluate 15 embedding models, ranging from…

Xinping Zhao, Xinshuo Hu, Jiaxin Xu, Danyu Tang, et al.
Published
Mar 2026
Upvotes
75
Citations
7
Read summary Original
63

arXiv.org

Project Imaging-X: A Survey of 1000+ Open-Access Medical Imaging Datasets for Foundation Model Development

This survey reviews over 1,000 open-access medical imaging datasets released between 2000 and 2025, analyzing their modalities, tasks, anatomical regions, and limitations. It finds the landscape is fragmented, small-scale, and unevenly distributed, with 2D images dominating, pathology and X-ray being the most common modalities, and classification and…

Zhongying Deng, Cheng Tang, Ziyan Huang, Jiashi Lin, et al.
Published
Mar 2026
Upvotes
72
Citations
2
Read summary Original
64

arXiv.org

MiroEval: Benchmarking Multimodal Deep Research Agents in Process and Outcome

MiroEval is a benchmark and evaluation framework for deep research systems, comprising 100 tasks (70 text-only, 30 multimodal) grounded in real user needs and built via a dual-path pipeline supporting periodic updates. The evaluation suite assesses systems along three complementary dimensions: adaptive synthesis quality with task-specific rubrics, agentic…

Fangda Ye, Yuxin Hu, Pengxiang Zhu, Yibo Li, et al.
Published
Mar 2026
Upvotes
72
Citations
4
Read summary Original
65

Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval

Learning to Retrieve from Agent Trajectories

This paper introduces LRAT (Learning to Retrieve from Agent Trajectories), a framework for training retrieval models directly from the interaction data of LLM-powered search agents. The authors argue that traditional human-centric retrieval training is misaligned with agentic search, where agents issue intermediate queries and consume results in multi-turn…

Yuqi Zhou, Sunhao Dai, Changle Qu, Liang Pang, et al.
Published
Mar 2026
Upvotes
72
Citations
2
Read summary Original
66

arXiv.org

Kinema4D: Kinematic 4D World Modeling for Spatiotemporal Embodied Simulation

Kinema4D is a new action-conditioned 4D generative robotic simulator that models robot-world interactions as 4D spatiotemporal events. It disentangles simulation into precise 4D robot control via URDF-based kinematics and generative 4D environmental reactions using a diffusion transformer that synthesizes synchronized RGB and pointmap sequences. The…

Mutian Xu, Tianbao Zhang, Tianqi Liu, Zhaoxi Chen, et al.
Published
Mar 2026
Upvotes
70
Citations
6
Read summary Original
67

NVIDIA

Nemotron-Cascade 2: Post-Training LLMs with Cascade RL and Multi-Domain On-Policy Distillation

Nemotron-Cascade 2 is an open 30B Mixture-of-Experts (MoE) model with 3B activated parameters, achieving best-in-class reasoning and agentic capabilities. It is the second open-weight LLM to achieve Gold Medal-level performance in the 2025 IMO, IOI, and ICPC World Finals, with 20x fewer parameters than DeepSeek-V3.2-Speciale-671B-A37B. The model builds on…

Zhuolin Yang, Zihan Liu, Yang Chen, Wenliang Dai, et al.
Published
Mar 2026
Upvotes
70
Citations
20
Read summary Original
68

arXiv.org

Calibri: Enhancing Diffusion Transformers via Parameter-Efficient Calibration

The paper introduces Calibri, a parameter-efficient method to enhance Diffusion Transformers (DiTs) by calibrating block outputs with learned scaling parameters. The authors show that selectively disabling or re-weighting DiT blocks can improve generation quality, leading to a black-box optimization problem solved via CMA-ES, modifying only ~10^2…

Danil Tokhchukov, Aysel Mirzoeva, Andrey Kuznetsov, Konstantin Sobolev
Published
Mar 2026
Upvotes
69
Citations
0
Read summary Original
69

arXiv.org

SAMA: Factorized Semantic Anchoring and Motion Alignment for Instruction-Guided Video Editing

SAMA is a framework for instruction-guided video editing that factorizes the task into semantic anchoring and motion modeling. It uses Semantic Anchoring to predict semantic tokens from sparse anchor frames, enabling instruction-aware structural planning, and Motion Alignment, which pre-trains the backbone on motion-centric pretext tasks (cube inpainting,…

Xinyao Zhang, Wenkai Dong, Yuxin Song, Bo Fang, et al.
Published
Mar 2026
Upvotes
68
Citations
3
Read summary Original
70

Google DeepMind

VGGRPO: Towards World-Consistent Video Generation with 4D Latent Reward

VGGRPO is a framework for geometry-aware post-training of video diffusion models, addressing geometric drift and unstable camera motion. It introduces a Latent Geometry Model (LGM) that stitches video diffusion latents to a geometry foundation model (e.g., Any4D) via a lightweight connector, enabling direct prediction of 4D scene geometry (camera poses,…

Zhaochong An, Orest Kupyn, Théo Uscidda, Andrea Colaco, et al.
Published
Mar 2026
Upvotes
68
Citations
14
Read summary Original
71

arXiv.org

Strategic Navigation or Stochastic Search? How Agents and Humans Reason Over Document Collections

The paper introduces MADQA, a benchmark of 2,250 human-authored questions over 800 heterogeneous PDF documents, designed to evaluate multimodal agentic systems. It formalizes Agentic Document Collection VQA with six properties: extractive answers, multi-hop reasoning, closed-world assumption, grounded attribution, agentic retrieval, and visual…

Łukasz Borchmann, Jordy Van Landeghem, Michał Turski, Shreyansh Padarha, et al.
Published
Mar 2026
Upvotes
65
Citations
1
Read summary Original
72

arXiv.org

Video-Oasis: Rethinking Evaluation of Video Understanding

The paper introduces Video-Oasis, a diagnostic suite for auditing video understanding benchmarks. It finds that 55% of samples in 14 existing benchmarks can be solved without visual or temporal context, and that state-of-the-art Video-LLMs perform only marginally above random chance on the remaining video-native challenges. The suite includes…

Geuntaek Lim, Sungjune Park, Jaeyun Lee, Inwoong Lee, et al.
Published
Mar 2026
Upvotes
65
Citations
1
Read summary Original
73

Annual Meeting of the Association for Computational Linguistics

RubricBench: Aligning Model-Generated Rubrics with Human Standards

RubricBench is a new benchmark with 1,147 pairwise comparisons designed to assess rubric-guided evaluation in reward models. It uses a multi-dimensional filtration pipeline to select hard samples with input complexity, output surface bias, and process failures, each annotated with human-derived atomic rubrics. Experiments show that while rubric-aware…

Qiyuan Zhang, Junyi Zhou, Yufei Wang, Fuyuan Lyu, et al.
Published
Mar 2026
Upvotes
64
Citations
10
Read summary Original
74

Z.ai / GLM

IndexCache: Accelerating Sparse Attention via Cross-Layer Index Reuse

IndexCache is a method to accelerate sparse attention in large language models by exploiting cross-layer redundancy in token selection. DeepSeek Sparse Attention (DSA) uses a lightning indexer at each layer to select top-k tokens, but this indexer still has O(L^2) complexity. IndexCache partitions layers into a small set of Full (F) layers that run their…

Yushi Bai, Qian Dong, Ting Jiang, Xin Lv, et al.
Published
Mar 2026
Upvotes
64
Citations
16
Read summary Original
75

Google DeepMind

LoGeR: Long-Context Geometric Reconstruction with Hybrid Memory

LoGeR (Long-context Geometric Reconstruction) is a novel architecture for scaling feedforward dense 3D reconstruction to extremely long video sequences (up to 19k frames) without post-optimization. It processes video in chunks, using bidirectional attention for intra-chunk reasoning and a hybrid memory module for inter-chunk coherence. The hybrid memory…

Junyi Zhang, Charles Herrmann, Junhwa Hur, Chen Sun, et al.
Published
Mar 2026
Upvotes
63
Citations
22
Read summary Original
76

Mistral AI

Voxtral TTS

Voxtral TTS is a multilingual zero-shot text-to-speech model that generates natural speech from as little as 3 seconds of reference audio. It uses a hybrid architecture: an autoregressive decoder backbone (based on Ministral 3B) predicts semantic speech tokens, while a flow-matching transformer predicts acoustic tokens. The tokens are produced by Voxtral…

Mistral-AI, :, Alexander H. Liu, Alexis Tacnet, et al.
Published
Mar 2026
Upvotes
63
Citations
0
Read summary Original
77

arXiv.org

SpecEyes: Accelerating Agentic Multimodal LLMs via Speculative Perception and Planning

SpecEyes is an agentic-level speculative acceleration framework for multimodal large language models (MLLMs) that addresses the sequential bottleneck of tool-use loops. It uses a lightweight, tool-free model to speculatively answer queries that do not require deep agentic reasoning, bypassing the expensive tool-calling pipeline. The framework consists of…

Haoyu Huang, Jinfa Huang, Zhongwei Wan, Xiawu Zheng, et al.
Published
Mar 2026
Upvotes
62
Citations
1
Read summary Original
78

arXiv.org

WorldCam: Interactive Autoregressive 3D Gaming Worlds with Camera Pose as a Unifying Geometric Representation

WorldCam is an interactive 3D gaming world model that uses camera pose as a unifying geometric representation to achieve precise action control, long-horizon generation, and 3D consistency. It maps user actions (keyboard/mouse) to 6-DoF camera poses via Lie algebra, injecting them into a video DiT (Wan-2.1) for accurate control. Global camera poses serve…

Jisu Nam, Yicong Hong, Chun-Hao Paul Huang, Feng Liu, et al.
Published
Mar 2026
Upvotes
61
Citations
6
Read summary Original
79

arXiv.org

How Far Can Unsupervised RLVR Scale LLM Training?

This paper investigates the scalability of Unsupervised Reinforcement Learning with Verifiable Rewards (URLVR) for LLM training. It categorizes URLVR methods into intrinsic (certainty-based and ensemble-based) and external rewards. The authors establish a unified theoretical framework showing that all intrinsic methods converge to sharpening the model's…

Bingxiang He, Yuxin Zuo, Zeyuan Liu, Shangziqi Zhao, et al.
Published
Mar 2026
Upvotes
60
Citations
24
Read summary Original
80

arXiv.org

Online Experiential Learning for Language Models

The paper introduces Online Experiential Learning (OEL), a framework for large language models to continuously improve from their own deployment experience without human annotations, reward models, or access to user-side environments. OEL operates in two iterated stages: first, transferable experiential knowledge is extracted and accumulated from user-side…

Tianzhu Ye, Li Dong, Qingxiu Dong, Xun Wu, et al.
Published
Mar 2026
Upvotes
60
Citations
20
Read summary Original
81

arXiv.org

FASTER: Rethinking Real-Time Flow VLAs

FASTER is a method to reduce reaction latency in flow-based Vision-Language-Action (VLA) models for real-time robot control. It introduces a Horizon-Aware Schedule (HAS) that adaptively allocates more sampling steps to near-term actions, enabling the immediate action to be generated in a single step while preserving long-horizon trajectory quality. This…

Yuxiang Lu, Zhe Liu, Xianzhe Fan, Zhenya Yang, et al.
Published
Mar 2026
Upvotes
60
Citations
10
Read summary Original
82

arXiv.org

TRUST-SQL: Tool-Integrated Multi-Turn Reinforcement Learning for Text-to-SQL over Unknown Schemas

TRUST-SQL addresses Text-to-SQL parsing under the Unknown Schema setting, where databases have hundreds of tables and noisy metadata, making full schema prefilling impractical. The task is formulated as a Partially Observable Markov Decision Process, and the agent follows a four-phase protocol (Explore, Propose, Generate, Confirm) to ground reasoning in…

Ai Jian, Xiaoyun Zhang, Wanrou Du, Jingqing Ruan, et al.
Published
Mar 2026
Upvotes
58
Citations
3
Read summary Original
83

arXiv.org

3DreamBooth: High-Fidelity 3D Subject-Driven Video Generation Model

The paper introduces 3DreamBooth, a framework for 3D-aware video customization that generates view-consistent videos of a subject from a few multi-view reference images. It addresses the limitation of existing subject-driven video generation methods that treat subjects as 2D entities, lacking 3D geometry priors. The framework comprises two components:…

Hyun-kyu Ko, Jihyeon Park, Younghyun Kim, Dongheok Park, et al.
Published
Mar 2026
Upvotes
58
Citations
0
Read summary Original
84

arXiv.org

Memento-Skills: Let Agents Design Agents

Memento-Skills is a generalist LLM agent system that functions as an agent-designing agent, autonomously constructing, adapting, and improving task-specific agents through experience. It uses a memory-based reinforcement learning framework with stateful prompts, where reusable skills stored as structured markdown files serve as persistent, evolving memory.…

Huichi Zhou, Siyuan Guo, Anjie Liu, Zhongwei Yu, et al.
Published
Mar 2026
Upvotes
58
Citations
39
Read summary Original
85

arXiv.org

RealRestorer: Towards Generalizable Real-World Image Restoration with Large-Scale Image Editing Models

RealRestorer is an open-source image restoration model designed to handle diverse real-world degradations, including blur, rain, noise, low-light, moiré patterns, haze, compression artifacts, reflection, and flare. The authors construct a large-scale dataset with a synthesis pipeline that combines synthetic and real-world degradation data, and they…

Yufeng Yang, Xianfang Zeng, Zhangqi Jiang, Fukun Yin, et al.
Published
Mar 2026
Upvotes
58
Citations
3
Read summary Original
86

arXiv.org

Gen-Searcher: Reinforcing Agentic Search for Image Generation

Gen-Searcher is the first trained multimodal deep search agent for image generation, addressing the limitation of frozen internal knowledge in text-to-image models. It performs multi-hop web search and reasoning to collect textual knowledge and reference images for grounded generation. The authors built a data pipeline producing two training datasets…

Kaituo Feng, Manyuan Zhang, Shuang Chen, Yunlong Lin, et al.
Published
Mar 2026
Upvotes
58
Citations
16
Read summary Original
87

arXiv.org

CHIMERA: Compact Synthetic Data for Generalizable LLM Reasoning

The paper introduces CHIMERA, a compact synthetic dataset of 9,225 samples designed to address data-centric challenges in LLM reasoning post-training: cold-start data scarcity, limited domain coverage, and the annotation bottleneck. The dataset is built via a three-stage pipeline: subject expansion using GPT-5 to create a hierarchical taxonomy of 1,179…

Xinyu Zhu, Yihao Feng, Yanchao Sun, Xianzhi Du, et al.
Published
Mar 2026
Upvotes
57
Citations
2
Read summary Original
88

arXiv.org

BeyondSWE: Can Current Code Agent Survive Beyond Single-Repo Bug Fixing?

BeyondSWE is a 500-instance benchmark from 246 real GitHub repositories that evaluates code agents beyond single-repository bug fixing, covering four settings: cross-repository issue resolution (CrossRepo), domain-specific issue resolution (DomainFix), dependency-driven migration (DepMigrate), and document-to-repository generation (Doc2Repo). These tasks…

Guoxin Chen, Fanzhe Meng, Jiale Zhao, Minghao Li, et al.
Published
Mar 2026
Upvotes
57
Citations
11
Read summary Original
89

arXiv.org

From Static Templates to Dynamic Runtime Graphs: A Survey of Workflow Optimization for LLM Agents

This survey reviews methods for optimizing workflows in large language model (LLM)-based agentic systems, which are modeled as agentic computation graphs (ACGs). The authors propose a taxonomy based on when workflow structure is determined, distinguishing static methods (fixed reusable templates optimized offline) from dynamic methods (structure selected,…

Ling Yue, Kushal Raj Bhandari, Ching-Yun Ko, Dhaval Patel, et al.
Published
Mar 2026
Upvotes
57
Citations
14
Read summary Original
90

arXiv.org

Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs?

This paper investigates why self-distillation can degrade the mathematical reasoning performance of LLMs, despite often improving other domains. The authors find that self-distillation suppresses 'epistemic verbalization'—the expression of uncertainty during reasoning—which is crucial for robust problem-solving. Through controlled experiments, they show…

Jeonghye Kim, Xufang Luo, Minbeom Kim, Sangmook Lee, et al.
Published
Mar 2026
Upvotes
57
Citations
61
Read summary Original
91

arXiv.org

BandPO: Bridging Trust Regions and Ratio Clipping via Probability-Aware Bounds for LLM Reinforcement Learning

The paper introduces BandPO, a method for LLM reinforcement learning that replaces the fixed clipping bounds of PPO/GRPO with dynamic, probability-aware bounds derived from f-divergence trust regions. The authors identify a bottleneck in canonical clipping: fixed bounds limit the upward update margin for low-probability actions, suppressing high-advantage…

Yuan Li, Bo Wang, Yufei Gao, Yuqian Yao, et al.
Published
Mar 2026
Upvotes
56
Citations
2
Read summary Original
92

arXiv.org

Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills

Trace2Skill is a framework that distills agent execution trajectories into portable skills by analyzing many traces in parallel and consolidating recurring lessons into a single skill directory. It supports both deepening existing human-written skills and creating skills from weak LLM-generated drafts. The pipeline has three stages: trajectory generation…

Jingwei Ni, Yihao Liu, Xinpeng Liu, Yutao Sun, et al.
Published
Mar 2026
Upvotes
56
Citations
78
Read summary Original
93

arXiv.org

DARE: Aligning LLM Agents with the R Statistical Ecosystem via Distribution-Aware Retrieval

The paper introduces DARE (Distribution-Aware Retrieval Embedding), a lightweight retrieval model that incorporates data distribution information into function representations for retrieving R packages, addressing the underuse of rigorous statistical methods in R by LLM agents. The authors construct RPKB, a curated knowledge base from 8,191 CRAN packages,…

Maojun Sun, Yue Wu, Yifei Xie, Ruijian Han, et al.
Published
Mar 2026
Upvotes
54
Citations
3
Read summary Original
94

NVIDIA

MM-Zero: Self-Evolving Multi-Model Vision Language Models From Zero Data

MM-Zero is a reinforcement learning framework that enables zero-data self-evolution for Vision Language Models (VLMs). It introduces a tri-role system—Proposer, Coder, and Solver—all initialized from the same base model and trained sequentially using Group Relative Policy Optimization (GRPO). The Proposer generates visual descriptions and questions, the…

Zongxia Li, Hongyang Du, Chengsong Huang, Xiyang Wu, et al.
Published
Mar 2026
Upvotes
54
Citations
12
Read summary Original
95

OpenAI

Reasoning Models Struggle to Control their Chains of Thought

The paper introduces CoT-Control, an evaluation suite with 14,076 instances, to measure CoT controllability—the ability of reasoning models to follow instructions that constrain their chain of thought (CoT). Across 13 frontier models, CoT controllability is significantly lower than output controllability (e.g., Claude Sonnet 4.5: 2.7% vs 61.9%).…

Chen Yueh-Han, Robert McCarthy, Bruce W. Lee, He He, et al.
Published
Mar 2026
Upvotes
39
Citations
11
Read summary Original
96

OpenAI

IH-Challenge: A Training Dataset to Improve Instruction Hierarchy on Frontier LLMs

The paper introduces IH-Challenge, a reinforcement learning (RL) training dataset designed to improve instruction hierarchy (IH) robustness in large language models (LLMs). IH defines how models prioritize system, developer, user, and tool instructions under conflict, which is key for defending against jailbreaks, system prompt extractions, and prompt…

Chuan Guo, Juan Felipe Ceron Uribe, Sicheng Zhu, Christopher A. Choquette-Choo, et al.
Published
Mar 2026
Upvotes
0
Citations
12
Read summary Original
97

Together AI

V1: Unifying Generation and Self-Verification for Parallel Reasoners

The paper introduces V1, a framework that unifies generation and self-verification for parallel reasoning in LLMs. It identifies that pointwise self-verification suffers from calibration collapse, while self-aggregation methods like RSA lead to diversity collapse. V1 comprises two components: V1-Infer, an uncertainty-guided pairwise ranking algorithm using…

Harman Singh, Xiuyu Li, Kusha Sareen, Monishwaran Maheswaran, et al.
Published
Mar 2026
Upvotes
Not indexed
Citations
7
Read summary Original
98

Together AI

Mamba-3: Improved Sequence Modeling using State Space Principles

Mamba-3 introduces three methodological improvements to state space models (SSMs) for better quality, capability, and inference efficiency. First, exponential-trapezoidal discretization generalizes prior heuristics, enabling a more expressive recurrence that can replace the short causal convolution. Second, complex-valued state transitions, implemented via…

Aakash Lahoti, Kevin Y. Li, Berlin Chen, Caitlin Wang, et al.
Published
Mar 2026
Upvotes
Not indexed
Citations
68
Read summary Original
99

Together AI

FlashAttention-4: Algorithm and Kernel Pipelining Co-Design for Asymmetric Hardware Scaling

FlashAttention-4 addresses the asymmetric hardware scaling of NVIDIA Blackwell GPUs (B200/GB200), where tensor core throughput doubles compared to Hopper, but shared memory bandwidth and exponential unit throughput do not scale accordingly. This shifts bottlenecks to non-matmul operations. The paper introduces techniques to mitigate these bottlenecks: (1)…

Ted Zadouri, Markus Hoehnerbach, Jay Shah, Timothy Liu, et al.
Published
Mar 2026
Upvotes
Not indexed
Citations
36
Read summary Original
100

Together AI

Speculative Speculative Decoding

The paper introduces speculative speculative decoding (SSD), a framework that parallelizes drafting and verification in speculative decoding (SD) by having the draft model predict likely verification outcomes and pre-speculate for them while verification runs. This eliminates drafting overhead on cache hits, making SSD lossless and faster than SD. The…

Tanishq Kumar, Tri Dao, Avner May
Published
Mar 2026
Upvotes
Not indexed
Citations
11
Read summary Original