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。
Yang Yue, Zhiqi Chen, Rui Lu +5 authors
Despite initial claims, reinforcement learning with verifiable rewards does not introduce fundamentally new reasoning abilities to LLMs, instead it enhances performance by biasing output distribution toward rewarded paths without expanding the reasoning boundary.
50 篇论文 · 按点赞排序
Minju Seo, Jinheon Baek, Seongyun Lee +1 authors
PaperCoder is a multi-agent LLM framework that converts machine learning papers into functional code repositories through planning, analysis, and generation stages.
Yuxin Zuo, Kaiyan Zhang, Shang Qu +7 authors
Test-Time Reinforcement Learning (TTRL) enhances Large Language Models (LLMs) using unlabeled data through reinforcement learning, improving performance across tasks.
Khalil Hennara, Sara Chrouf, Mohamed Motaism Hamed +3 authors
A method for integrating a new language into a large language model enhances performance in the target language while preserving existing knowledge.
Shiyu Liu, Yucheng Han, Peng Xing +21 authors
Step1X-Edit, an image editing model that uses multimodal LLM and diffusion image decoder, outperforms open-source models and approaches proprietary models in quality.
Jianhao Yan, Yafu Li, Zican Hu +5 authors
LUFFY enhances zero-RL models with off-policy guidance, improving reasoning and generalization through balanced imitation and exploration.
Weiye Xu, Jiahao Wang, Weiyun Wang +10 authors
VisuLogic is a benchmark to evaluate genuine vision-centric reasoning in multimodal large language models, revealing significant performance gaps compared to human accuracy.
Guo Chen, Zhiqi Li, Shihao Wang +18 authors
Eagle 2.5 improves long-context multimodal learning through Automatic Degrade Sampling and Image Area Preservation techniques, enhancing VLMs for video and image understanding, and matches state-of-the-art performance on benchmarks.
Long Lian, Yifan Ding, Yunhao Ge +8 authors
The Describe Anything Model (DAM) leverages a focal prompt and localized vision backbone to achieve detailed localized captioning, outperforming existing models on various benchmarks through a semi-supervised data pipeline.
Minghao Wu, Weixuan Wang, Sinuo Liu +7 authors
Research reveals significant disparities in multilingual benchmark evaluations, emphasizing the need for culturally and linguistically tailored benchmarks over translations to achieve equitable technological progress.
Shangshang Wang, Julian Asilis, Ömer Faruk Akgül +3 authors
Tina, a family of tiny reasoning models, achieves high reasoning performance at minimal computational cost through LoRA during reinforcement learning.
Aviv Slobodkin, Hagai Taitelbaum, Yonatan Bitton +8 authors
RefVNLI, a novel metric, evaluates both textual alignment and subject preservation in subject-driven text-to-image generation, outperforming existing methods across various benchmarks and subject categories.
Fulong Ye, Miao Hua, Pengze Zhang +5 authors
A diffusion-based face swapping model, DreamID, achieves high identity similarity, attribute preservation, and fast inference by using explicit Triplet ID Group supervision and an improved diffusion architecture.
Cheng Qian, Emre Can Acikgoz, Qi He +5 authors
A comprehensive study on reward design for tool use within reinforcement learning improves LLMs' tool use capabilities and generalization performance over supervised fine-tuning.
Hongcheng Gao, Yue Liu, Yufei He +6 authors
A meta-agent named FlowReasoner automates the design of query-level multi-agent systems using DeepSeek R1 and reinforcement learning, excelling in performance, complexity, and efficiency across benchmarks.
Jiayi Pan, Xiuyu Li, Long Lian +6 authors
Adaptive Parallel Reasoning (APR) enhances language model performance by optimally combining serialized and parallel computations through adaptive multi-threading and reinforcement learning.
Tianyang Xu, Haojie Zheng, Chengze Li +4 authors
NodeRAG, a graph-centric framework with heterogeneous graph structures, enhances Retrieval-augmented Generation (RAG) by improving integration and performance in indexing, querying, and question-answering.
Tiancheng Gu, Kaicheng Yang, Ziyong Feng +6 authors
A two-stage framework using Multimodal Large Language Models learns discriminative embeddings for various tasks by distilling knowledge from a powerful LLM and enhancing instruction tuning with hard negatives.
Yicheng Chen, Yining Li, Kai Hu +3 authors
A unified method quantifies dataset information content using a label graph and an efficient MIG sampling technique to enhance diversity and quality in instruction-tuning datasets, outperforming existing methods.
Joya Chen, Ziyun Zeng, Yiqi Lin +3 authors
Large-scale training for Video LLMs using ASR transcripts enables competitive video QA performance and real-time commentary capabilities.
Shi Qiu, Shaoyang Guo, Zhuo-Yang Song +49 authors
A new benchmark evaluates large language models on physics problems with a novel Expression Edit Distance metric, revealing gaps compared to human experts.
Salman Rahman, Liwei Jiang, James Shiffer +7 authors
X-Teaming is a scalable framework that explores and generates multi-turn attack scenarios against language models, achieving high success rates, and XGuard-Train is a large multi-turn safety training dataset to mitigate such attacks.
Sungjun Han, Juyoung Suk, Suyeong An +5 authors
Trillion-7B is a highly efficient multilingual LLM leveraging Cross-lingual Document Attention (XLDA) for knowledge transfer and achieving competitive performance with minimal multilingual training data.
Shaden Alshammari, John Hershey, Axel Feldmann +2 authors
A unified framework using KL divergence between supervisory and learned representations generalizes multiple machine learning loss functions and improves unsupervised image classification and debiasing.
Phillip Y. Lee, Jihyeon Je, Chanho Park +3 authors
A framework named Abstract Perspective Change (APC) enhances perspective-aware reasoning in vision-language models by utilizing vision foundation models to create scene abstractions and enable perspective shifts.
Yiting Ran, Xintao Wang, Tian Qiu +3 authors
BookWorld, a system for simulating multi-agent societies based on books, generates high-quality, faithful stories and supports applications like story generation and social simulation.
Chaoyun Zhang, He Huang, Chiming Ni +18 authors
UFO2, an AgentOS for Windows, integrates multimodal LLMs with native APIs and hybrid control detection to enhance robust and scalable desktop automation.
Changjiang Gao, Xu Huang, Wenhao Zhu +3 authors
Multilingual reasoning in large language models shows significant potential for improved performance in reasoning tasks compared to English-only models.
Minho Park, Taewoong Kang, Jooyeol Yun +2 authors
SphereDiff uses spherical latent representations to generate high-fidelity 360-degree panoramic content by reducing the distortions of equirectangular projection and leveraging pretrained diffusion models.
Ivan Moshkov, Darragh Hanley, Ivan Sorokin +5 authors
This paper presents our winning submission to the AI Mathematical Olympiad - Progress Prize 2 (AIMO-2) competition. Our recipe for building state-of-the-art mathematical reasoning models relies on three key pillars. First, we create a large-scale dataset comprising 540K unique high-quality math problems, including olympiad-level problems, and their 3.2M long-reasoning solutions. Second, we develop a novel method to integrate code execution with long reasoning models through iterative training, generation, and quality filtering, resulting in 1.7M high-quality Tool-Integrated Reasoning solutions. Third, we create a pipeline to train models to select the most promising solution from many candidates. We show that such generative solution selection (GenSelect) can significantly improve upon majority voting baseline. Combining these ideas, we train a series of models that achieve state-of-the-art results on mathematical reasoning benchmarks. To facilitate further research, we release our code, models, and the complete OpenMathReasoning dataset under a commercially permissive license.
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