The Year in AI Papers
The year/January 2026

Monthly collection

January 2026

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

Papers
86
Citations
1.8K
Research labs
8

Explore the year

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

86 papers published in January 2026

01

arXiv.org

Green-VLA: Staged Vision-Language-Action Model for Generalist Robots

Green-VLA is a staged Vision-Language-Action (VLA) framework for real-world robot deployment, developed by Sber Robotics Center. It uses a five-stage curriculum: L0 (base VLM), L1 (web pretraining), R0 (multi-embodiment robotics pretraining), R1 (embodiment-specific fine-tuning), and R2 (RL alignment). The framework unifies 24M web samples and 3,000 hours…

I. Apanasevich, M. Artemyev, R. Babakyan, P. Fedotova, et al.
Published
Jan 2026
Upvotes
322
Citations
4
Read summary Original
02

NVIDIA

GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization

The paper identifies a flaw in applying Group Relative Policy Optimization (GRPO) to multi-reward reinforcement learning: normalizing the summed reward causes distinct reward combinations to collapse into identical advantage values, reducing training signal resolution and causing suboptimal convergence or early failure. To address this, the authors propose…

Shih-Yang Liu, Xin Dong, Ximing Lu, Shizhe Diao, et al.
Published
Jan 2026
Upvotes
235
Citations
114
Read summary Original
03

Google DeepMind

PaperBanana: Automating Academic Illustration for AI Scientists

PaperBanana is an agentic framework that automates the generation of publication-ready academic illustrations, such as methodology diagrams and statistical plots. It uses five specialized agents—Retriever, Planner, Stylist, Visualizer, and Critic—powered by VLMs and image generation models. The Retriever selects relevant reference examples, the Planner…

Dawei Zhu, Rui Meng, Yale Song, Xiyu Wei, et al.
Published
Jan 2026
Upvotes
229
Citations
19
Read summary Original
04

arXiv.org

Watching, Reasoning, and Searching: A Video Deep Research Benchmark on Open Web for Agentic Video Reasoning

The paper introduces VideoDR, the first benchmark for video deep research, which requires models to extract multi-frame visual anchors from videos, perform interactive web searches, and conduct multi-hop reasoning over combined video-web evidence to answer open-domain factoid questions. The benchmark comprises 500 samples across six semantic domains, with…

Chengwen Liu, Xiaomin Yu, Zhuoyue Chang, Zhe Huang, et al.
Published
Jan 2026
Upvotes
215
Citations
6
Read summary Original
05

arXiv.org

Agentic Reasoning for Large Language Models

This survey introduces agentic reasoning, a paradigm shift where large language models (LLMs) act as autonomous agents that plan, act, and learn through continual interaction with their environment, rather than passively generating sequences. The authors organize agentic reasoning along three complementary dimensions: foundational agentic reasoning (core…

Tianxin Wei, Ting-Wei Li, Zhining Liu, Xuying Ning, et al.
Published
Jan 2026
Upvotes
207
Citations
34
Read summary Original
06

arXiv.org

BabyVision: Visual Reasoning Beyond Language

The paper introduces BABYVISION, a benchmark to evaluate core visual abilities in Multimodal LLMs (MLLMs) that are independent of linguistic knowledge, targeting skills humans develop before language. It contains 388 questions across 22 subtypes in four categories: Fine-grained Discrimination, Visual Tracking, Spatial Perception, and Visual Pattern…

Liang Chen, Weichu Xie, Yiyan Liang, Hongfeng He, et al.
Published
Jan 2026
Upvotes
201
Citations
24
Read summary Original
07

arXiv.org

STEP3-VL-10B Technical Report

Step3-VL-10B is a 10B-parameter open-source multimodal foundation model that rivals or surpasses models 10-20x larger, such as GLM-4.6V-106B and Qwen3-VL-235B, and proprietary systems like Gemini 2.5 Pro. It achieves 92.2% on MMBench, 80.11% on MMMU, 94.43% on AIME2025, and 75.95% on MathVision. The model uses a unified pre-training strategy on 1.2T…

Ailin Huang, Chengyuan Yao, Chunrui Han, Fanqi Wan, et al.
Published
Jan 2026
Upvotes
196
Citations
26
Read summary Original
08

arXiv.org

Can LLMs Clean Up Your Mess? A Survey of Application-Ready Data Preparation with LLMs

This paper surveys the use of large language models (LLMs) for data preparation, covering data cleaning, integration, and enrichment. It contrasts traditional rule-based and model-specific methods with LLM-enhanced approaches that leverage prompting, retrieval-augmented generation (RAG), fine-tuning, and agentic workflows. The survey identifies three core…

Wei Zhou, Jun Zhou, Haoyu Wang, Zhenghao Li, et al.
Published
Jan 2026
Upvotes
191
Citations
6
Read summary Original
09

arXiv.org

LTX-2: Efficient Joint Audio-Visual Foundation Model

LTX-2 is an open-source text-to-audio+video (T2AV) foundation model that jointly generates synchronized video and audio from text. It uses an asymmetric dual-stream transformer with a 14B-parameter video stream and a 5B-parameter audio stream, connected via bidirectional cross-attention layers with temporal positional embeddings and cross-modality AdaLN.…

Yoav HaCohen, Benny Brazowski, Nisan Chiprut, Yaki Bitterman, et al.
Published
Jan 2026
Upvotes
185
Citations
135
Read summary Original
10

arXiv.org

LongCat-Flash-Thinking-2601 Technical Report

LongCat-Flash-Thinking-2601 is a 560B-parameter open-source Mixture-of-Experts (MoE) reasoning model with 27B activated parameters, achieving state-of-the-art performance among open-source models on agentic benchmarks. Its training combines a unified framework with domain-parallel expert training, environment scaling, and a robust RL pipeline. Key…

Meituan LongCat Team, Anchun Gui, Bei Li, Bingyang Tao, et al.
Published
Jan 2026
Upvotes
183
Citations
23
Read summary Original
11

arXiv.org

Idea2Story: An Automated Pipeline for Transforming Research Concepts into Complete Scientific Narratives

Idea2Story is a framework for autonomous scientific discovery that shifts literature understanding from online reasoning to offline knowledge construction. It builds a structured methodological knowledge graph from peer-reviewed papers (about 13,000 from NeurIPS and ICLR over three years) and their review feedback, extracting reusable method units and…

Tengyue Xu, Zhuoyang Qian, Gaoge Liu, Li Ling, et al.
Published
Jan 2026
Upvotes
183
Citations
5
Read summary Original
12

Annual Meeting of the Association for Computational Linguistics

Thinking with Map: Reinforced Parallel Map-Augmented Agent for Geolocalization

The paper introduces Thinking with Map, a map-augmented agent for image geolocalization that equips a large vision-language model (LVLM) with map tools (POI search, static/satellite map queries, image zoom) to iteratively propose and verify location hypotheses. The method uses a two-stage optimization: agentic reinforcement learning (GRPO) to improve…

Yuxiang Ji, Yong Wang, Ziyu Ma, Yiming Hu, et al.
Published
Jan 2026
Upvotes
171
Citations
8
Read summary Original
13

arXiv.org

Your Group-Relative Advantage Is Biased

This paper identifies a fundamental bias in group-relative advantage estimation used by RLVR algorithms like GRPO. The authors prove that this estimator systematically underestimates advantages for hard prompts and overestimates them for easy prompts, with bias increasing as prompt difficulty deviates from 0.5 and with smaller group sizes. To address this,…

Fengkai Yang, Zherui Chen, Xiaohan Wang, Xiaodong Lu, et al.
Published
Jan 2026
Upvotes
158
Citations
27
Read summary Original
14

arXiv.org

Urban Socio-Semantic Segmentation with Vision-Language Reasoning

The paper introduces urban socio-semantic segmentation, targeting entities defined by social attributes (e.g., schools, parks) rather than physical ones. The authors present SocioSeg, a benchmark with over 13,000 samples, organizing labels into three hierarchical tasks: socio-name, socio-class, and socio-function. A key innovation is representing…

Yu Wang, Yi Wang, Rui Dai, Yujie Wang, et al.
Published
Jan 2026
Upvotes
155
Citations
2
Read summary Original
15

arXiv.org

Vision-DeepResearch: Incentivizing DeepResearch Capability in Multimodal Large Language Models

Vision-DeepResearch introduces a new paradigm for multimodal deep research, addressing the hit-rate problem in image search and the limited reasoning depth and search breadth of existing methods. The approach enables multi-turn, multi-entity, and multi-scale visual and textual search, supporting dozens of reasoning steps and hundreds of engine…

Wenxuan Huang, Yu Zeng, Qiuchen Wang, Zhen Fang, et al.
Published
Jan 2026
Upvotes
155
Citations
25
Read summary Original
16

Annual Meeting of the Association for Computational Linguistics

Rewarding the Rare: Uniqueness-Aware RL for Creative Problem Solving in LLMs

The paper introduces Uniqueness-Aware Reinforcement Learning (UARL) to address exploration collapse in RL-trained LLMs, where policies prematurely converge to a few dominant reasoning patterns, improving pass@1 but limiting pass@k. UARL operates at the rollout level, using an LLM judge to cluster solutions by high-level strategy and reweighting GRPO…

Zhiyuan Hu, Yucheng Wang, Yufei He, Jiaying Wu, et al.
Published
Jan 2026
Upvotes
151
Citations
9
Read summary Original
17

arXiv.org

Advancing Open-source World Models

LingBot-World is an open-source world simulator built from video generation, offering high-fidelity dynamics across diverse environments, minute-level long-term memory, and real-time interactivity with sub-second latency. It uses a three-stage training pipeline: pre-training on a 14B-parameter Wan2.2 model, middle-training with a mixture-of-experts (MoE)…

Robbyant Team, Zelin Gao, Qiuyu Wang, Yanhong Zeng, et al.
Published
Jan 2026
Upvotes
135
Citations
87
Read summary Original
18

arXiv.org

NeoVerse: Enhancing 4D World Model with in-the-wild Monocular Videos

NeoVerse is a 4D world model that reconstructs dynamic 4D Gaussian Splatting (4DGS) from monocular videos in a feed-forward, pose-free manner, enabling novel-trajectory video generation and downstream applications. It addresses scalability limitations of prior methods by avoiding expensive multi-view data and offline preprocessing. Key innovations include…

Yuxue Yang, Lue Fan, Ziqi Shi, Junran Peng, et al.
Published
Jan 2026
Upvotes
133
Citations
28
Read summary Original
19

arXiv.org

DeepResearchEval: An Automated Framework for Deep Research Task Construction and Agentic Evaluation

DeepResearchEval is an automated framework for constructing deep research tasks and evaluating deep research systems. It addresses limitations in existing benchmarks: annotation-intensive task construction, static evaluation dimensions, and incomplete fact verification. The framework uses a persona-driven pipeline to generate realistic, complex tasks…

Yibo Wang, Lei Wang, Yue Deng, Keming Wu, et al.
Published
Jan 2026
Upvotes
128
Citations
12
Read summary Original
20

arXiv.org

daVinci-Dev: Agent-native Mid-training for Software Engineering

The paper introduces daVinci-Dev, a training recipe for agentic software engineering that uses agent-native mid-training data to bridge the gap between static training corpora and dynamic, feedback-rich development environments. The authors construct two complementary trajectory types: contextually-native trajectories (68.6B tokens) derived from GitHub…

Ji Zeng, Dayuan Fu, Tiantian Mi, Yumin Zhuang, et al.
Published
Jan 2026
Upvotes
126
Citations
10
Read summary Original
21

Research paper

AgentDoG: A Diagnostic Guardrail Framework for AI Agent Safety and Security

AgentDoG is a diagnostic guardrail framework for AI agent safety and security, developed by the Shanghai Artificial Intelligence Laboratory. It addresses limitations in existing guardrails by introducing a unified three-dimensional safety taxonomy that categorizes agentic risks by source (where), failure mode (how), and real-world harm (what). Guided by…

Dongrui Liu, Qihan Ren, Chen Qian, Shuai Shao, et al.
Published
Jan 2026
Upvotes
126
Citations
24
Read summary Original
22

arXiv.org

Harder Is Better: Boosting Mathematical Reasoning via Difficulty-Aware GRPO and Multi-Aspect Question Reformulation

The paper introduces MathForge, a framework to enhance mathematical reasoning in large language models via reinforcement learning with verifiable rewards (RLVR). It addresses a lack of emphasis on harder questions in existing methods, both algorithmically and data-wise. Algorithmically, it proposes Difficulty-Aware Group Policy Optimization (DGPO), which…

Yanqi Dai, Yuxiang Ji, Xiao Zhang, Yong Wang, et al.
Published
Jan 2026
Upvotes
119
Citations
23
Read summary Original
23

arXiv.org

Entropy-Adaptive Fine-Tuning: Resolving Confident Conflicts to Mitigate Forgetting

The paper investigates why Supervised Fine-Tuning (SFT) causes catastrophic forgetting while on-policy Reinforcement Learning (RL) preserves general capabilities. The authors identify a distributional gap: SFT data contains 'Confident Conflicts'—tokens with low probability and low entropy, where the model is confident but forced to learn a divergent label,…

Muxi Diao, Lele Yang, Wuxuan Gong, Yutong Zhang, et al.
Published
Jan 2026
Upvotes
115
Citations
12
Read summary Original
24

arXiv.org

Controlled Self-Evolution for Algorithmic Code Optimization

The paper introduces Controlled Self-Evolution (CSE), a framework to improve the exploration efficiency of self-evolution methods for algorithmic code optimization. CSE addresses three bottlenecks: initialization bias, uncontrolled stochastic evolution, and insufficient experience utilization. It comprises three components: Diversified Planning…

Tu Hu, Ronghao Chen, Shuo Zhang, Jianghao Yin, et al.
Published
Jan 2026
Upvotes
115
Citations
11
Read summary Original
25

NVIDIA

Golden Goose: A Simple Trick to Synthesize Unlimited RLVR Tasks from Unverifiable Internet Text

The paper introduces Golden Goose, a method to synthesize unlimited Reinforcement Learning with Verifiable Rewards (RLVR) tasks from unverifiable internet text by converting them into multiple-choice fill-in-the-middle questions. Given a source text, an LLM masks a contiguous span of crucial reasoning steps (the ground-truth answer) and generates diverse,…

Ximing Lu, David Acuna, Jaehun Jung, Jian Hu, et al.
Published
Jan 2026
Upvotes
113
Citations
6
Read summary Original
26

arXiv.org

Everything in Its Place: Benchmarking Spatial Intelligence of Text-to-Image Models

The paper introduces SpatialGenEval, a benchmark for evaluating the spatial intelligence of text-to-image (T2I) models. It uses 1,230 long, information-dense prompts across 25 real-world scenes, each integrating 10 spatial sub-domains (object, attribute, position, orientation, layout, comparison, proximity, occlusion, motion, causal) and paired with 10…

Zengbin Wang, Xuecai Hu, Yong Wang, Feng Xiong, et al.
Published
Jan 2026
Upvotes
111
Citations
8
Read summary Original
27

arXiv.org

MMFormalizer: Multimodal Autoformalization in the Wild

The paper introduces MMFORMALIZER, a framework for multimodal autoformalization that translates natural language and visual inputs into formal LEAN statements. It addresses challenges in grounding physical concepts by using recursive grounding and dimensional analysis, with adaptive termination to ensure visual evidence supports abstractions. The authors…

Jing Xiong, Qi Han, Yunta Hsieh, Hui Shen, et al.
Published
Jan 2026
Upvotes
106
Citations
0
Read summary Original
28

arXiv.org

Scaling Embeddings Outperforms Scaling Experts in Language Models

This technical report from Meituan's LongCat team investigates scaling embeddings as an alternative to scaling experts in Mixture-of-Experts (MoE) language models. The authors find that N-gram Embedding, which augments token embeddings with hashed n-gram sub-tables, achieves a superior Pareto frontier compared to increasing expert count in specific…

Hong Liu, Jiaqi Zhang, Chao Wang, Xing Hu, et al.
Published
Jan 2026
Upvotes
106
Citations
11
Read summary Original
29

arXiv.org

InfiniDepth: Arbitrary-Resolution and Fine-Grained Depth Estimation with Neural Implicit Fields

InfiniDepth introduces a novel depth representation that models depth as neural implicit fields, enabling arbitrary-resolution and fine-grained monocular depth estimation. The method uses a Vision Transformer encoder to extract multi-scale features, which are queried at continuous 2D coordinates via bilinear interpolation and fused hierarchically through a…

Hao Yu, Haotong Lin, Jiawei Wang, Jiaxin Li, et al.
Published
Jan 2026
Upvotes
104
Citations
12
Read summary Original
30

Research paper

K-EXAONE Technical Report

K-EXAONE is a 236B-parameter Mixture-of-Experts (MoE) language model developed by LG AI Research, activating 23B parameters per token. It supports a 256K-token context window and six languages: Korean, English, Spanish, German, Japanese, and Vietnamese. The model uses a hybrid attention mechanism (global and sliding window) and a fine-grained sparse MoE…

Eunbi Choi, Kibong Choi, Seokhee Hong, Junwon Hwang, et al.
Published
Jan 2026
Upvotes
97
Citations
2
Read summary Original
31

Annual Meeting of the Association for Computational Linguistics

MAXS: Meta-Adaptive Exploration with LLM Agents

The paper introduces MAXS, a meta-adaptive reasoning framework for LLM agents that addresses two issues: locally myopic generation and trajectory instability. MAXS uses a lookahead strategy to extend reasoning paths a few steps ahead, estimating the advantage value of tool usage, and combines step consistency variance and inter-step trend slopes to select…

Jian Zhang, Zhiyuan Wang, Zhangqi Wang, Yu He, et al.
Published
Jan 2026
Upvotes
96
Citations
1
Read summary Original
32

arXiv.org

SWE-Pruner: Self-Adaptive Context Pruning for Coding Agents

SWE-Pruner is a self-adaptive context pruning framework for coding agents, addressing the high API costs and latency from long interaction contexts. It uses a lightweight 0.6B neural skimmer, trained on 61K synthetic samples, to perform task-aware, line-level pruning based on a goal hint provided by the agent. Evaluated on multi-turn tasks (SWE-Bench…

Yuhang Wang, Yuling Shi, Mo Yang, Rongrui Zhang, et al.
Published
Jan 2026
Upvotes
94
Citations
26
Read summary Original
33

arXiv.org

Collaborative Multi-Agent Test-Time Reinforcement Learning for Reasoning

MATTRL is a framework that improves multi-agent LLM reasoning at test time by injecting structured textual experience into multi-agent deliberation, avoiding costly weight updates. It forms a team of specialist agents, retrieves relevant experiences from a pool, and reaches consensus. The experience pool is built by scoring utterances with individual and…

Zhiyuan Hu, Yunhai Hu, Juncheng Liu, Shuyue Stella Li, et al.
Published
Jan 2026
Upvotes
92
Citations
3
Read summary Original
34

arXiv.org

EvoCUA: Evolving Computer Use Agents via Learning from Scalable Synthetic Experience

EvoCUA is a native computer-use agent that replaces static imitation with a self-sustaining evolutionary cycle integrating verifiable data synthesis, scalable interaction infrastructure, and iterative policy optimization. A synthesis engine generates diverse tasks with executable validators, while a high-throughput sandbox platform orchestrates tens of…

Taofeng Xue, Chong Peng, Mianqiu Huang, Linsen Guo, et al.
Published
Jan 2026
Upvotes
92
Citations
31
Read summary Original
35

arXiv.org

Evolving Programmatic Skill Networks

The paper introduces the Programmatic Skill Network (PSN), a framework for continual skill acquisition in open-ended embodied environments. In PSN, skills are executable symbolic programs (e.g., JavaScript for Minecraft) that form a compositional network evolving through experience. PSN implements three core mechanisms via LLMs: REFLECT for trace-based…

Haochen Shi, Xingdi Yuan, Bang Liu
Published
Jan 2026
Upvotes
88
Citations
4
Read summary Original
36

Annual Meeting of the Association for Computational Linguistics

PaCoRe: Learning to Scale Test-Time Compute with Parallel Coordinated Reasoning

PaCoRe (Parallel Coordinated Reasoning) is a training-and-inference framework that scales test-time compute (TTC) beyond sequential reasoning limits by using parallel exploration coordinated via message passing. It runs multiple rounds: each round generates parallel reasoning trajectories, compacts them into context-bounded messages, and synthesizes these…

Jingcheng Hu, Yinmin Zhang, Shijie Shang, Xiaobo Yang, et al.
Published
Jan 2026
Upvotes
87
Citations
16
Read summary Original
37

Research paper

Computer Environments Elicit General Agentic Intelligence in LLMs

The paper introduces LLM-in-Sandbox, a paradigm that virtualizes a computer as a minimal code sandbox (with bash, file_editor, and finish tools) to elicit general agentic intelligence in LLMs. Without additional training, strong models (e.g., Claude-Sonnet-4.5-Think, GPT-5, DeepSeek-V3.2-Thinking) show consistent gains across mathematics, physics,…

Daixuan Cheng, Shaohan Huang, Yuxian Gu, Huatong Song, et al.
Published
Jan 2026
Upvotes
87
Citations
5
Read summary Original
38

arXiv.org

Innovator-VL: A Multimodal Large Language Model for Scientific Discovery

Innovator-VL is a scientific multimodal large language model (MLLM) designed for scientific understanding and reasoning while maintaining general vision performance. It uses a transparent, reproducible pipeline with RICE-ViT vision encoder, PatchMerger projector, and Qwen3-8B-Base language model. Training includes language-image alignment (LLaVA-1.5 558k),…

Zichen Wen, Boxue Yang, Shuang Chen, Yaojie Zhang, et al.
Published
Jan 2026
Upvotes
85
Citations
10
Read summary Original
39

Volume 1

CAR-bench: Evaluating the Consistency and Limit-Awareness of LLM Agents under Real-World Uncertainty

CAR-bench is a benchmark for evaluating LLM agents in real-world, user-facing applications, specifically in-car voice assistants. It addresses limitations of existing benchmarks that overlook reliability under uncertainty. The environment includes an LLM-simulated user, 19 domain policies, 58 interconnected tools, and dynamic states/databases. Beyond…

Johannes Kirmayr, Lukas Stappen, Elisabeth André
Published
Jan 2026
Upvotes
85
Citations
6
Read summary Original
40

arXiv.org

$A^3$-Bench: Benchmarking Memory-Driven Scientific Reasoning via Anchor and Attractor Activation

The paper introduces A3-Bench, a benchmark for evaluating memory-driven scientific reasoning in large language models (LLMs). It is grounded in the concepts of anchors (foundational knowledge units) and attractors (experience-based templates), which are activated during reasoning. The authors annotate 2,198 science problems across math, physics, and…

Jian Zhang, Yu He, Zhiyuan Wang, Zhangqi Wang, et al.
Published
Jan 2026
Upvotes
84
Citations
1
Read summary Original
41

arXiv.org

MemGovern: Enhancing Code Agents through Learning from Governed Human Experiences

MemGovern is a framework that transforms raw GitHub issue-tracking data into structured, agent-friendly experiential memory to enhance autonomous software engineering (SWE) agents. It addresses the 'closed-world' limitation of agents that ignore historical human debugging experience. MemGovern uses experience governance to filter, standardize, and…

Qihao Wang, Ziming Cheng, Shuo Zhang, Fan Liu, et al.
Published
Jan 2026
Upvotes
82
Citations
8
Read summary Original
42

arXiv.org

FlowAct-R1: Towards Interactive Humanoid Video Generation

FlowAct-R1 is a framework for real-time interactive humanoid video generation, built on a MMDiT architecture (Seedance). It enables streaming synthesis of arbitrary-length videos with low latency, achieving 25fps at 480p resolution with a time-to-first-frame of about 1.5 seconds. The method uses a chunkwise diffusion forcing strategy with a self-forcing…

Lizhen Wang, Yongming Zhu, Zhipeng Ge, Youwei Zheng, et al.
Published
Jan 2026
Upvotes
78
Citations
7
Read summary Original
43

arXiv.org

Being-H0.5: Scaling Human-Centric Robot Learning for Cross-Embodiment Generalization

Being-H0.5 is a foundational Vision-Language-Action (VLA) model for cross-embodiment generalization, introduced by the BeingBeyond Team. It uses a human-centric learning paradigm, treating human interaction traces as a universal 'mother tongue' for physical interaction. The model is trained on UniHand-2.0, the largest embodied pre-training recipe to date,…

Hao Luo, Ye Wang, Wanpeng Zhang, Sipeng Zheng, et al.
Published
Jan 2026
Upvotes
77
Citations
53
Read summary Original
44

Qwen

Qwen3-TTS Technical Report

The Qwen3-TTS technical report introduces a family of multilingual, controllable, robust, and streaming text-to-speech models. Trained on over 5 million hours of speech data across 10 languages, Qwen3-TTS supports 3-second voice cloning, description-based voice design, and fine-grained control. It uses a dual-track LM architecture with two tokenizers:…

Hangrui Hu, Xinfa Zhu, Ting He, Dake Guo, et al.
Published
Jan 2026
Upvotes
77
Citations
88
Read summary Original
45

arXiv.org

DynamicVLA: A Vision-Language-Action Model for Dynamic Object Manipulation

DynamicVLA is a 0.4B-parameter Vision-Language-Action (VLA) model designed for dynamic object manipulation, addressing latency issues in existing VLAs. It introduces three key innovations: a compact architecture using a convolutional vision encoder (FastViT) and a truncated SmolLM2-360M backbone for fast inference; Continuous Inference, which overlaps…

Haozhe Xie, Beichen Wen, Jiarui Zheng, Zhaoxi Chen, et al.
Published
Jan 2026
Upvotes
76
Citations
24
Read summary Original
46

Annual Meeting of the Association for Computational Linguistics

HERMES: KV Cache as Hierarchical Memory for Efficient Streaming Video Understanding

HERMES is a training-free framework for efficient streaming video understanding in Multimodal Large Language Models (MLLMs). It conceptualizes the KV cache as a hierarchical memory system based on a mechanistic attention analysis, where shallow layers act as sensory memory with recency bias, deep layers as long-term memory focusing on frame-level anchor…

Haowei Zhang, Shudong Yang, Jinlan Fu, See-Kiong Ng, et al.
Published
Jan 2026
Upvotes
75
Citations
22
Read summary Original
47

Research paper

The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models

This paper challenges the assumption that arbitrary-order generation in Diffusion Large Language Models (dLLMs) enhances reasoning potential. The authors find that for general reasoning tasks like mathematics and coding, arbitrary-order generation can actually limit reasoning potential compared to standard autoregressive (AR) order. They introduce the…

Zanlin Ni, Shenzhi Wang, Yang Yue, Tianyu Yu, et al.
Published
Jan 2026
Upvotes
75
Citations
2
Read summary Original
48

DeepSeek

DeepSeek-OCR 2: Visual Causal Flow

DeepSeek-OCR 2 introduces DeepEncoder V2, a novel vision encoder that replaces the CLIP component with a compact LLM (Qwen2-0.5B) to enable causal reordering of visual tokens, mimicking human visual scanning. The encoder uses a dual attention mask: bidirectional for visual tokens and causal for learnable query tokens, allowing queries to attend to all…

Haoran Wei, Yaofeng Sun, Yukun Li
Published
Jan 2026
Upvotes
73
Citations
53
Read summary Original
49

NVIDIA

Motion Attribution for Video Generation

The paper introduces Motive, a gradient-based data attribution framework for video generation models that isolates motion from static appearance. It uses motion-weighted loss masks, computed via AllTracker optical flow, to focus influence scores on dynamic regions. Motive scales via single-timestep estimation, common randomness, frame-length normalization,…

Xindi Wu, Despoina Paschalidou, Jun Gao, Antonio Torralba, et al.
Published
Jan 2026
Upvotes
72
Citations
2
Read summary Original
50

Research paper

Solar Open Technical Report

Solar Open is a 102B-parameter bilingual Mixture-of-Experts language model developed by Upstage to address data scarcity for underserved languages, focusing on Korean. The model uses a 196,608-token BPE tokenizer, a sparse MoE architecture with 12B active parameters, and was trained on 20 trillion tokens. Key innovations include synthesizing 4.5T tokens of…

Sungrae Park, Sanghoon Kim, Jungho Cho, Gyoungjin Gim, et al.
Published
Jan 2026
Upvotes
67
Citations
2
Read summary Original
51

Annual Meeting of the Association for Computational Linguistics

ABC-Bench: Benchmarking Agentic Backend Coding in Real-World Development

ABC-Bench is a benchmark for evaluating LLM-based agents on full-lifecycle backend development tasks, covering repository exploration, code editing, environment configuration, deployment, and end-to-end API testing. It includes 224 tasks from 127 MIT-licensed GitHub repositories, spanning 8 languages and 19 frameworks. The ABC-Pipeline automates task…

Jie Yang, Honglin Guo, Li Ji, Jiazheng Zhou, et al.
Published
Jan 2026
Upvotes
67
Citations
3
Read summary Original
52

arXiv.org

MOSS Transcribe Diarize Technical Report

MOSS Transcribe Diarize is a unified multimodal large language model for Speaker-Attributed, Time-Stamped Transcription (SATS), jointly performing word recognition, speaker attribution, and timestamp prediction in a single end-to-end pass. It uses a 128k-token context window to process up to 90 minutes of audio without chunking, preserving long-range…

MOSI. AI, :, Donghua Yu, Zhengyuan Lin, et al.
Published
Jan 2026
Upvotes
65
Citations
5
Read summary Original
53

arXiv.org

NextFlow: Unified Sequential Modeling Activates Multimodal Understanding and Generation

NextFlow is a unified decoder-only autoregressive transformer trained on 6 trillion interleaved text-image tokens. It uses a dual-codebook tokenizer for semantic and pixel-level features, and adopts next-scale prediction for visual generation instead of raster-scan, enabling 1024x1024 image generation in 5 seconds. The model retains next-token prediction…

Huichao Zhang, Liao Qu, Yiheng Liu, Hang Chen, et al.
Published
Jan 2026
Upvotes
64
Citations
12
Read summary Original
54

arXiv.org

VIBE: Visual Instruction Based Editor

VIBE is a compact, high-throughput instruction-based image editing pipeline that combines a 2B-parameter Qwen3-VL model for instruction interpretation with a 1.6B-parameter Sana1.5 diffusion model for image generation. The architecture uses channel-wise concatenation for reference image guidance and learnable meta-tokens processed by a lightweight…

Grigorii Alekseenko, Aleksandr Gordeev, Irina Tolstykh, Bulat Suleimanov, et al.
Published
Jan 2026
Upvotes
64
Citations
2
Read summary Original
55

Mistral AI

Ministral 3

The Ministral 3 series introduces three parameter-efficient dense language models (3B, 8B, and 14B) derived from the Mistral Small 3.1 24B parent model via Cascade Distillation, an iterative pruning and distillation technique. Each size has base, instruct, and reasoning variants, all with vision capabilities and Apache 2.0 licensing. The models support up…

Alexander H. Liu, Kartik Khandelwal, Sandeep Subramanian, Victor Jouault, et al.
Published
Jan 2026
Upvotes
64
Citations
53
Read summary Original
56

arXiv.org

Distribution-Aligned Sequence Distillation for Superior Long-CoT Reasoning

This report introduces DASD-4B-Thinking, a lightweight open-source reasoning model that achieves state-of-the-art performance on math, science, and code benchmarks, outperforming larger models (e.g., 32B) using only 448K training samples. The authors critique the common sequence-level distillation paradigm (SFT on teacher-generated responses) for three…

Shaotian Yan, Kaiyuan Liu, Chen Shen, Bing Wang, et al.
Published
Jan 2026
Upvotes
64
Citations
13
Read summary Original
57

arXiv.org

Advances and Frontiers of LLM-based Issue Resolution in Software Engineering: A Comprehensive Survey

This paper presents a systematic survey of LLM-based issue resolution in software engineering, a task formalized by benchmarks like SWE-bench. The survey reviews 175 papers, organizing the field into a taxonomy covering data, methods, and analysis. Data is categorized into evaluation and training datasets, with construction via automated collection or…

Caihua Li, Lianghong Guo, Yanlin Wang, Daya Guo, et al.
Published
Jan 2026
Upvotes
63
Citations
7
Read summary Original
58

arXiv.org

Quartet II: Accurate LLM Pre-Training in NVFP4 by Improved Unbiased Gradient Estimation

The paper introduces Quartet II, a fully-NVFP4 quantization scheme for LLM pre-training on NVIDIA Blackwell GPUs. It proposes MS-EDEN, a new unbiased quantization routine for microscaling formats that reduces quantization error by more than 2x compared to stochastic rounding (SR), by applying randomized Hadamard rotations and merging EDEN correction…

Andrei Panferov, Erik Schultheis, Soroush Tabesh, Dan Alistarh
Published
Jan 2026
Upvotes
63
Citations
11
Read summary Original
59

Annual Meeting of the Association for Computational Linguistics

RubricHub: A Comprehensive and Highly Discriminative Rubric Dataset via Automated Coarse-to-Fine Generation

The paper introduces RubricHub, a large-scale (~110k) multi-domain rubric dataset, generated via an automated Coarse-to-Fine framework to address limitations in existing rubric-based evaluation for open-ended tasks. The framework comprises three stages: principle-guided and response-grounded generation, multi-model aggregation, and difficulty evolution,…

Sunzhu Li, Jiale Zhao, Miteto Wei, Huimin Ren, et al.
Published
Jan 2026
Upvotes
62
Citations
29
Read summary Original
60

arXiv.org

MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods

MMFineReason is a large-scale multimodal reasoning dataset with 1.8M samples and 5.1B solution tokens, built via a three-stage pipeline: data aggregation and standardization, CoT distillation from Qwen3-VL-235B-A22B-Thinking, and quality/difficulty-based selection. The dataset covers mathematics (79.4%), science (13.8%), puzzle/game (4.6%), and general/OCR…

Honglin Lin, Zheng Liu, Yun Zhu, Chonghan Qin, et al.
Published
Jan 2026
Upvotes
62
Citations
33
Read summary Original
61

arXiv.org

TwinBrainVLA: Unleashing the Potential of Generalist VLMs for Embodied Tasks via Asymmetric Mixture-of-Transformers

TwinBrainVLA is a novel Vision-Language-Action (VLA) framework that addresses catastrophic forgetting in VLM fine-tuning for robotic control. It uses two isomorphic VLM pathways: a frozen 'Left Brain' (generalist) preserving pre-trained semantic knowledge, and a trainable 'Right Brain' (specialist) for motor control. An Asymmetric Mixture-of-Transformers…

Bin Yu, Shijie Lian, Xiaopeng Lin, Yuliang Wei, et al.
Published
Jan 2026
Upvotes
61
Citations
12
Read summary Original
62

arXiv.org

ASTRA: Automated Synthesis of agentic Trajectories and Reinforcement Arenas

ASTRA is a fully automated, end-to-end framework for training tool-augmented language model agents. It combines a trajectory synthesis pipeline that uses the static topology of tool-call graphs to generate diverse, multi-turn tool-use trajectories for supervised fine-tuning (SFT), with an environment synthesis framework that converts decomposed…

Xiaoyu Tian, Haotian Wang, Shuaiting Chen, Hao Zhou, et al.
Published
Jan 2026
Upvotes
61
Citations
3
Read summary Original
63

arXiv.org

The Molecular Structure of Thought: Mapping the Topology of Long Chain-of-Thought Reasoning

The paper proposes that effective long chain-of-thought (Long CoT) reasoning in LLMs requires stable molecular-like structures formed by three interaction types: Deep-Reasoning (covalent-like), Self-Reflection (hydrogen-bond-like), and Self-Exploration (van der Waals-like). Distillation from strong reasoning LLMs (e.g., DeepSeek-R1, QwQ, OpenAI-OSS)…

Qiguang Chen, Yantao Du, Ziniu Li, Jinhao Liu, et al.
Published
Jan 2026
Upvotes
60
Citations
3
Read summary Original
64

Qwen

Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking

The report introduces Qwen3-VL-Embedding and Qwen3-VL-Reranker, a unified framework for multimodal retrieval built on the Qwen3-VL foundation model. The embedding model uses a multi-stage training pipeline (contrastive pre-training, multi-task contrastive learning, and reranker distillation) to produce high-dimensional vectors, supporting Matryoshka…

Mingxin Li, Yanzhao Zhang, Dingkun Long, Keqin Chen, et al.
Published
Jan 2026
Upvotes
59
Citations
167
Read summary Original
65

Annual Meeting of the Association for Computational Linguistics

KnowMe-Bench: Benchmarking Person Understanding for Lifelong Digital Companions

KnowMe-Bench is a benchmark for evaluating person understanding in lifelong digital companions, built from long-form autobiographical narratives rather than sparse chat logs. It addresses two gaps in existing benchmarks: evaluation misalignment (retrieval proxies vs. true person understanding) and data substrate misalignment (low-density, decontextualized…

Tingyu Wu, Zhisheng Chen, Ziyan Weng, Shuhe Wang, et al.
Published
Jan 2026
Upvotes
59
Citations
8
Read summary Original
66

arXiv.org

Avatar Forcing: Real-Time Interactive Head Avatar Generation for Natural Conversation

Avatar Forcing is a framework for real-time interactive head avatar generation that models user-avatar interactions using diffusion forcing. It addresses two key challenges: real-time motion generation under causal constraints and learning expressive reactions without labeled data. The framework processes multimodal user inputs (audio and motion) with low…

Taekyung Ki, Sangwon Jang, Jaehyeong Jo, Jaehong Yoon, et al.
Published
Jan 2026
Upvotes
58
Citations
13
Read summary Original
67

arXiv.org

Toward Efficient Agents: Memory, Tool learning, and Planning

This survey examines efficiency in LLM-based agents, focusing on memory, tool use, and planning. It defines an efficient agent as one that maximizes task success while minimizing resource consumption (tokens, latency, computational cost). The survey reviews methods for efficient memory (construction, management, access, skills, multi-agent), tool use…

Xiaofang Yang, Lijun Li, Heng Zhou, Tong Zhu, et al.
Published
Jan 2026
Upvotes
57
Citations
9
Read summary Original
68

NVIDIA

Fast-ThinkAct: Efficient Vision-Language-Action Reasoning via Verbalizable Latent Planning

Fast-ThinkAct is an efficient reasoning framework for Vision-Language-Action (VLA) tasks that compresses lengthy chain-of-thought (CoT) reasoning into compact latent representations. It uses a teacher-student distillation approach, where a textual teacher VLM (trained with GRPO) provides reasoning traces, and a latent student VLM learns to generate…

Chi-Pin Huang, Yunze Man, Zhiding Yu, Min-Hung Chen, et al.
Published
Jan 2026
Upvotes
56
Citations
17
Read summary Original
69

arXiv.org

Stable-DiffCoder: Pushing the Frontier of Code Diffusion Large Language Model

Stable-DiffCoder is a diffusion-based language model for code that reuses the Seed-Coder architecture, data, and training pipeline but replaces autoregressive (AR) training with a block diffusion continual pretraining (CPT) stage. The authors introduce a tailored warmup and a block-wise clipped noise schedule to stabilize training and improve knowledge…

Chenghao Fan, Wen Heng, Bo Li, Sichen Liu, et al.
Published
Jan 2026
Upvotes
56
Citations
13
Read summary Original
70

arXiv.org

The Script is All You Need: An Agentic Framework for Long-Horizon Dialogue-to-Cinematic Video Generation

This paper introduces an agentic framework for generating long-horizon cinematic videos from dialogue. The framework comprises three agents: ScripterAgent, which translates dialogue into detailed cinematic scripts; DirectorAgent, which orchestrates video generation using a cross-scene continuous generation strategy with frame-anchoring to ensure temporal…

Chenyu Mu, Xin He, Qu Yang, Wanshun Chen, et al.
Published
Jan 2026
Upvotes
56
Citations
3
Read summary Original
71

arXiv.org

Scaling Text-to-Image Diffusion Transformers with Representation Autoencoders

This paper investigates scaling Representation Autoencoders (RAEs) for large-scale text-to-image (T2I) generation. The authors train RAE decoders on a frozen SigLIP-2 encoder using web, synthetic, and text-rendering data, finding that data composition is crucial for text reconstruction. They show that dimension-dependent noise scheduling remains essential,…

Shengbang Tong, Boyang Zheng, Ziteng Wang, Bingda Tang, et al.
Published
Jan 2026
Upvotes
55
Citations
40
Read summary Original
72

International Conference on 3D Vision

CaricatureGS: Exaggerating 3D Gaussian Splatting Faces With Gaussian Curvature

CaricatureGS introduces a method for creating photorealistic, controllable 3D caricature avatars by combining curvature-based geometric deformation with 3D Gaussian Splatting (3DGS). The pipeline starts with a multiview video, extracts a FLAME mesh, and solves a curvature-weighted Poisson equation to produce an exaggerated mesh. To train the 3DGS,…

Eldad Matmon, Amit Bracha, Noam Rotstein, Ron Kimmel
Published
Jan 2026
Upvotes
54
Citations
0
Read summary Original
73

Qwen

ArenaRL: Scaling RL for Open-Ended Agents via Tournament-based Relative Ranking

ArenaRL is a reinforcement learning framework for open-ended LLM agents that replaces pointwise scalar rewards with tournament-based relative ranking to overcome discriminative collapse, where reward models fail to distinguish subtle differences among high-quality trajectories. It introduces a process-aware pairwise evaluation mechanism and a seeded…

Qiang Zhang, Boli Chen, Fanrui Zhang, Ruixue Ding, et al.
Published
Jan 2026
Upvotes
54
Citations
9
Read summary Original
74

arXiv.org

LangForce: Bayesian Decomposition of Vision Language Action Models via Latent Action Queries

LangForce is a framework for Vision-Language-Action (VLA) models that addresses the 'vision shortcut' pathology, where models ignore language instructions due to dataset bias. In goal-driven datasets, language is predictable from vision, causing conditional mutual information between instructions and actions to vanish (Information Collapse). LangForce uses…

Shijie Lian, Bin Yu, Xiaopeng Lin, Laurence T. Yang, et al.
Published
Jan 2026
Upvotes
54
Citations
14
Read summary Original
75

arXiv.org

Why Can't I Open My Drawer? Mitigating Object-Driven Shortcuts in Zero-Shot Compositional Action Recognition

This paper addresses object-driven shortcuts in Zero-Shot Compositional Action Recognition (ZS-CAR), where models predict verbs by relying on object class labels rather than temporal evidence. The authors identify two root causes: sparse and skewed compositional supervision, and asymmetric learning difficulty (objects are easier to recognize than verbs).…

Geo Ahn, Inwoong Lee, Taeoh Kim, Minho Shim, et al.
Published
Jan 2026
Upvotes
54
Citations
0
Read summary Original
76

arXiv.org

DreamID-V:Bridging the Image-to-Video Gap for High-Fidelity Face Swapping via Diffusion Transformer

The paper introduces DreamID-V, a Diffusion Transformer (DiT)-based framework for high-fidelity video face swapping (VFS). It addresses the gap between image face swapping (IFS) and VFS by proposing a data pipeline, SyncID-Pipe, which pre-trains an Identity-Anchored Video Synthesizer (IVS) to generate synthetic videos, combined with IFS models to create…

Xu Guo, Fulong Ye, Xinghui Li, Pengqi Tu, et al.
Published
Jan 2026
Upvotes
53
Citations
7
Read summary Original
77

arXiv.org

MHLA: Restoring Expressivity of Linear Attention via Token-Level Multi-Head

The paper introduces Multi-Head Linear Attention (MHLA), a linear attention mechanism that restores expressivity by partitioning tokens into blocks (heads) along the token dimension, computing local key-value summaries, and mixing them with query-specific learned coefficients. This addresses 'global context collapse' in standard linear attention, which…

Kewei Zhang, Ye Huang, Yufan Deng, Jincheng Yu, et al.
Published
Jan 2026
Upvotes
53
Citations
5
Read summary Original
78

arXiv.org

User-Oriented Multi-Turn Dialogue Generation with Tool Use at scale

This paper introduces a user-oriented simulation framework for generating multi-turn tool-use dialogue data at scale, addressing limitations of static toolsets and single-shot trajectories. The authors first developed a task-oriented pipeline using an LRM-based simulator to dynamically synthesize tools and tasks, but found it produced minimal-interaction,…

Jungho Cho, Minbyul Jeong, Sungrae Park
Published
Jan 2026
Upvotes
53
Citations
1
Read summary Original
79

Volume 1

Paper2Rebuttal: A Multi-Agent Framework for Transparent Author Response Assistance

The paper introduces REBUTTALAGENT, a multi-agent framework for assisting authors in writing peer-review rebuttals. It reframes rebuttal generation as an evidence-centric planning task, decomposing feedback into atomic concerns, constructing hybrid contexts from compressed and raw manuscript text, and integrating on-demand external literature search. The…

Qianli Ma, Chang Guo, Zhiheng Tian, Siyu Wang, et al.
Published
Jan 2026
Upvotes
53
Citations
6
Read summary Original
80

Anthropic

How AI Impacts Skill Formation

This study examines how AI assistance affects skill formation in software engineering. In a randomized experiment, 52 developers learned a new asynchronous Python library (Trio) with or without an AI assistant. Results show that AI use significantly reduced quiz scores measuring conceptual understanding, code reading, and debugging (17% lower, Cohen's…

Judy Hanwen Shen, Alex Tamkin
Published
Jan 2026
Upvotes
10
Citations
24
Read summary Original
81

Moonshot AI

WorldVQA: Measuring Atomic World Knowledge in Multimodal Large Language Models

WorldVQA is a benchmark introduced to evaluate the atomic visual world knowledge of Multimodal Large Language Models (MLLMs), decoupling visual knowledge retrieval from reasoning. It comprises 3,500 VQA pairs across nine semantic categories, from common head-class entities to long-tail rarities. The benchmark follows four design principles: atomic…

Runjie Zhou, Youbo Shao, Haoyu Lu, Bowei Xing, et al.
Published
Jan 2026
Upvotes
6
Citations
4
Read summary Original
82

Anthropic

Constitutional Classifiers++: Efficient Production-Grade Defenses against Universal Jailbreaks

This paper introduces enhanced Constitutional Classifiers, a production-grade defense system against universal jailbreaks for large language models. The authors identify vulnerabilities in previous-generation defenses, such as reconstruction and output obfuscation attacks, and address them with exchange classifiers that evaluate outputs in full…

Hoagy Cunningham, Jerry Wei, Zihan Wang, Andrew Persic, et al.
Published
Jan 2026
Upvotes
0
Citations
28
Read summary Original
83

DeepSeek

Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models

The paper introduces Engram, a conditional memory module that augments Mixture-of-Experts (MoE) large language models with static N-gram lookup tables, providing a complementary sparsity axis to conditional computation. Engram uses hashed N-gram embeddings with tokenizer compression, multi-head hashing, and context-aware gating to retrieve static knowledge…

Xin Cheng, Rui Tian, Wangding Zeng, Damai Dai, et al.
Published
Jan 2026
Upvotes
0
Citations
63
Read summary Original
84

Anthropic

Who's in Charge? Disempowerment Patterns in Real-World LLM Usage

This paper presents the first large-scale empirical analysis of disempowerment patterns in real-world AI assistant interactions, analyzing 1.5 million consumer Claude.ai conversations using a privacy-preserving approach. The authors define situational disempowerment as occurring when interactions risk leading users to distorted perceptions of reality,…

Mrinank Sharma, Miles McCain, Raymond Douglas, David Duvenaud
Published
Jan 2026
Upvotes
0
Citations
19
Read summary Original
85

Together AI

DSGym: A Holistic Framework for Evaluating and Training Data Science Agents

DSGym is a standardized framework for evaluating and training data science agents in isolated, stateful execution environments. It addresses limitations in existing benchmarks, which often allow tasks to be solved without accessing data files (shortcut solvability), lack cross-benchmark standardization, and have narrow domain coverage. DSGym provides a…

Fan Nie, Junlin Wang, Harper Hua, Federico Bianchi, et al.
Published
Jan 2026
Upvotes
Not indexed
Citations
12
Read summary Original
86

Together AI

Learning to Discover at Test Time

The paper introduces TTT-Discover, a method that performs reinforcement learning at test time to solve scientific discovery problems. Unlike prior test-time scaling methods that use a frozen LLM for search, TTT-Discover continues to train the LLM on the specific test problem, using an entropic objective and PUCT-based state reuse to prioritize the most…

Mert Yuksekgonul, Daniel Koceja, Xinhao Li, Federico Bianchi, et al.
Published
Jan 2026
Upvotes
Not indexed
Citations
71
Read summary Original