Explore · 每周精选
发现最受关注的研究论文,追踪研究趋势,订阅感兴趣的期刊与关键词。
Shelly Bensal, Umar Jamil, Christopher Bryant +5 authors
A method using self-reflection and reinforcement learning improves the performance of large language models, especially with limited feedback, by rewarding self-reflections that lead to better task performance.
50 篇论文 · 按点赞排序
Shenzhi Wang, Le Yu, Chang Gao +15 authors
Token entropy patterns are crucial in RLVR, with high-entropy tokens significantly impacting reasoning performance and model optimization.
Mustafa Shukor, Dana Aubakirova, Francesco Capuano +11 authors
SmolVLA is a compact, efficient vision-language-action model that achieves competitive performance at reduced computational costs and can be deployed on consumer-grade hardware.
Mingjie Liu, Shizhe Diao, Ximing Lu +5 authors
Prolonged reinforcement learning training (ProRL) uncovers novel reasoning strategies in language models, outperforming base models and suggesting meaningful expansion of reasoning capabilities.
Junyu Zhang, Runpei Dong, Han Wang +8 authors
AlphaOne dynamically modulates reasoning in large models by introducing $\alpha$ moment and Bernoulli process for slow thinking, improving efficiency and capability across diverse domains.
Yanzhao Zhang, Mingxin Li, Dingkun Long +9 authors
The Qwen3 Embedding series, built on Qwen3 foundation models, offers advanced text embedding and reranking capabilities through a multi-stage training pipeline, achieving state-of-the-art performance across multilingual and retrieval benchmarks.
Zhenran Xu, Xue Yang, Yiyu Wang +7 authors
ComfyUI-Copilot uses a large language model and multi-agent system to enhance the usability and efficiency of the AI-driven art creation platform ComfyUI.
Ujjwal Upadhyay, Mukul Ranjan, Zhiqiang Shen +1 authors
SpookyBench is a benchmark for temporal pattern recognition in videos that highlights the limitations of vision-language models in processing noise-like frames without spatial information.
Xiaomi LLM-Core Team, Zihao Yue, Zhenru Lin +71 authors
MiMo-VL-7B-SFT and MiMo-VL-7B-RL provide state-of-the-art general visual understanding and multimodal reasoning through four-stage pre-training and Mixed On-policy Reinforcement Learning, outperforming models with up to 78B parameters.
Zafir Stojanovski, Oliver Stanley, Joe Sharratt +4 authors
Reasoning Gym provides a library of reasoning environments with verifiable rewards and procedural data generation for reinforcement learning, enabling the evaluation and training of reasoning models at varying difficulty levels.
Nikhil Kandpal, Brian Lester, Colin Raffel +24 authors
The Common Pile v0.1 dataset, an 8-terabyte collection of openly licensed text, is used to train competitive 7 billion parameter LLMs.
Jianyi Wang, Shanchuan Lin, Zhijie Lin +10 authors
SeedVR2 is a one-step diffusion-based video restoration model that uses adaptive window attention and feature matching loss to achieve high visual quality with reduced computational cost compared to existing methods.
Bin Lin, Zongjian Li, Xinhua Cheng +9 authors
A unified generative framework called UniWorld uses semantic features from visual-language models for image perception and manipulation, outperforming BAGEL with reduced data.
Zelai Xu, Zhexuan Xu, Xiangmin Yi +5 authors
VS-Bench is a multimodal benchmark designed to evaluate Vision Language Models' strategic reasoning and decision-making in complex multi-agent environments.
Etash Guha, Ryan Marten, Sedrick Keh +47 authors
The OpenThoughts project created open-source datasets leading to reasoning models that match or exceed state-of-the-art benchmarks in math, code, and science.
Tong Wu, Shuai Yang, Ryan Po +4 authors
A new framework enhances video world models' long-term consistency by integrating a geometry-grounded long-term spatial memory mechanism.
Qianhui Wu, Kanzhi Cheng, Rui Yang +15 authors
GUI-Actor, a VLM-based method using attention for coordinate-free GUI grounding, outperforms existing methods with better generalization and efficient fine-tuning.
Zijian Wu, Jinjie Ni, Xiangyan Liu +3 authors
SynthRL, a scalable pipeline for data synthesis in reinforcement learning with verifiable rewards, enhances visual math reasoning VLMs by generating challenging, verifiable questions.
Yihong Tang, Menglin Kong, Lijun Sun
LLMSynthor enhances LLMs for efficient and statistically accurate data synthesis through distributional feedback and proposal sampling.
Ai Jian, Weijie Qiu, Xiaokun Wang +6 authors
A new benchmark, CSVQA, evaluates scientific reasoning in vision-language models through domain-specific visual question answering, highlighting the need for improvement in these models.
Shuang Chen, Yue Guo, Zhaochen Su +7 authors
Inspired by the remarkable reasoning capabilities of Deepseek-R1 in complex textual tasks, many works attempt to incentivize similar capabilities in Multimodal Large Language Models (MLLMs) by directly applying reinforcement learning (RL). However, they still struggle to activate complex reasoning. In this paper, rather than examining multimodal RL in isolation, we delve into current training pipelines and identify three crucial phenomena: 1) Effective cold start initialization is critical for enhancing MLLM reasoning. Intriguingly, we find that initializing with carefully selected text data alone can lead to performance surpassing many recent multimodal reasoning models, even before multimodal RL. 2) Standard GRPO applied to multimodal RL suffers from gradient stagnation, which degrades training stability and performance. 3) Subsequent text-only RL training, following the multimodal RL phase, further enhances multimodal reasoning. This staged training approach effectively balances perceptual grounding and cognitive reasoning development. By incorporating the above insights and addressing multimodal RL issues, we introduce ReVisual-R1, achieving a new state-of-the-art among open-source 7B MLLMs on challenging benchmarks including MathVerse, MathVision, WeMath, LogicVista, DynaMath, and challenging AIME2024 and AIME2025.
Zhongwei Wan, Zhihao Dou, Che Liu +10 authors
A two-stage reinforcement learning framework enhances reasoning and self-reflection in multimodal large language models through reflection-focused datasets and a novel reward mechanism.
Anastasiia Ivanova, Eva Bakaeva, Zoya Volovikova +2 authors
AmbiK, a textual dataset of ambiguous instructions for kitchen robots, enables unified comparison of ambiguity detection methods.
Zhongmou He, Yee Man Choi, Kexun Zhang +6 authors
HARDTESTGEN creates a large, high-quality competitive programming dataset to enhance the precision and recall of verifiers in evaluating LLM-generated code.
Enshen Zhou, Jingkun An, Cheng Chi +8 authors
RoboRefer, a 3D-aware vision language model, enhances spatial understanding and multi-step reasoning in embodied robots through supervised and reinforcement fine-tuning, using the RefSpatial dataset and RefSpatial-Bench benchmark.
Mengdi Jia, Zekun Qi, Shaochen Zhang +5 authors
A comprehensive benchmark called OmniSpatial evaluates vision-language models' understanding of advanced spatial reasoning tasks, revealing significant limitations across various models.
Kinam Kim, Junha Hyung, Jaegul Choo
Temporal In-Context Fine-Tuning (TIC-FT) enhances pretrained video diffusion models for diverse conditional generation tasks with minimal data and without architectural changes.
Siyuan Li, Juanxi Tian, Zedong Wang +4 authors
SGG, an optimizer wrapper, enhances adaptive learning rates for large language models by grouping gradients and applying cluster-specific scaling, improving convergence and stability.
Ahmed Heakl, Sarim Hashmi, Gustavo Bertolo Stahl +3 authors
CASS is a dataset and model suite for GPU code transpilation at both source and assembly levels, achieving high accuracy and performance matching with native code.
Danil Sivtsov, Ivan Rodkin, Gleb Kuzmin +2 authors
Diagonal Batching enables parallel inference in Recurrent Memory Transformers, significantly improving speed and efficiency for long-context tasks.
北京市昌平区探索星信息技术及软件开发工作室
京ICP备2026059466号