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。
50 篇论文 · 按点赞排序
Wenjun Li, Zhi Chen, Jingru Lin +8 authors
Reinforcement learning is explored as a foundational approach for training deep research systems, addressing limitations of supervised and preference alignment methods by optimizing policies for tool interaction and exploration.
Zetong Zhou, Dongping Chen, Zixian Ma +6 authors
ReVPT enhances multi-modal LLMs' visual reasoning capabilities using reinforcement learning, achieving state-of-the-art performance on visual benchmarks.
Yufeng Cheng, Wenxu Wu, Shaojin Wu +3 authors
UMO, a Unified Multi-identity Optimization framework, enhances identity consistency and reduces confusion in multi-reference image customization using reinforcement learning on diffusion models.
Runpeng Dai, Linfeng Song, Haolin Liu +8 authors
Curiosity-Driven Exploration (CDE) enhances Reinforcement Learning with Verifiable Rewards (RLVR) by using intrinsic curiosity signals from the actor and critic to improve exploration and reduce premature convergence in Large Language Models (LLMs).
Changfeng Ma, Yang Li, Xinhao Yan +7 authors
P3-SAM, a native 3D point-promptable part segmentation model, achieves precise and robust segmentation of complex 3D objects using a feature extractor, multiple segmentation heads, and an IoU predictor.
Touayouch Brahim, Fosse Loïc, Damnati Géraldine +1 authors
Multi-task learning (MTL) is often achieved by merging datasets before fine-tuning, but the growing availability of fine-tuned models has led to new approaches such as model merging via task arithmetic. A major challenge in this setting is task interference, which worsens as the number of tasks increases. We propose a method that merges models trained on different tasks into a single model, maintaining strong performance across all tasks. Our approach leverages Jensen-Shannon divergence to guide the merging process without requiring additional labelled data, and automatically balances task importance. Unlike existing methods, our approach remains robust as the number of tasks grows and consistently outperforms prior work.
Zhuoqing Song, Peng Sun, Huizhuo Yuan +1 authors
CASTLE, an attention mechanism that updates keys with future context while maintaining autoregressive properties, outperforms standard causal attention in language modeling.
Sidharth Surapaneni, Hoang Nguyen, Jash Mehta +5 authors
AU-Harness is an efficient and comprehensive evaluation framework for Large Audio Language Models (LALMs) that addresses issues of speed, reproducibility, and task coverage, revealing gaps in temporal understanding and spoken language reasoning.
Yuyao Ge, Shenghua Liu, Yiwei Wang +6 authors
Contrastive Attention Refinement for Visual Enhancement (CARVE) improves VLM performance by extracting task-relevant visual signals through attention contrasting, addressing issues with visual complexity and attention mechanisms.
Xiangwei Shen, Zhimin Li, Zhantao Yang +6 authors
Direct-Align and Semantic Relative Preference Optimization improve diffusion models' alignment with human preferences by reducing computational costs and minimizing offline reward adaptation.
Ruofan Liang, Kai He, Zan Gojcic +4 authors
LuxDiT, a video diffusion transformer fine-tuned with low-rank adaptation, generates accurate HDR environment maps from visual input, outperforming existing methods.
Xuan-Phi Nguyen, Shrey Pandit, Revanth Gangi Reddy +4 authors
Continual reinforcement learning enhances autonomous single-agent models for deep research tasks, improving agentic skills and reasoning ability using synthetic data.
Wenxuan Huang, Shuang Chen, Zheyong Xie +15 authors
Interleaving Reasoning Generation (IRG) framework alternates between text-based thinking and image synthesis to improve Text-to-Image generation, achieving state-of-the-art performance and enhanced visual quality.
Lukas Haas, Gal Yona, Giovanni D'Antonio +2 authors
SimpleQA Verified is a refined benchmark for evaluating the factuality of Large Language Models, addressing issues in previous benchmarks and providing a more reliable evaluation tool.
Wenting Zhao, Pranjal Aggarwal, Swarnadeep Saha +3 authors
A reinforcement learning approach to aggregating multiple solutions for large language models improves performance on reasoning tasks by learning to synthesize correct answers from candidate solutions.
Mao Zheng, Zheng Li, Bingxin Qu +4 authors
Hunyuan-MT-7B and Hunyuan-MT-Chimera-7B are multilingual translation models that outperform existing models, especially in translating between Mandarin and minority languages, through a combination of pre-training, supervised fine-tuning, and reinforcement learning.
Duomin Wang, Wei Zuo, Aojie Li +7 authors
UniVerse-1, a unified audio-video generation model, uses a stitching of experts technique to combine pre-trained video and music models, ensuring accurate temporal alignment and producing high-quality audio-visual outputs.
Sergey Pletenev, Daniil Moskovskiy, Alexander Panchenko
Models fine-tuned on synthetic toxic data generated by LLMs perform worse than those trained on human data due to a lexical diversity gap in the synthetic content.
Ran Xin, Zeyu Zheng, Yanchen Nie +2 authors
BFS-Prover-V2 addresses scaling challenges in automated theorem proving by integrating a multi-turn off-policy RL framework and a planner-enhanced multi-agent search architecture, achieving state-of-the-art results on formal mathematics benchmarks.
Gagan Mundada, Yash Vishe, Amit Namburi +4 authors
WildScore evaluates MLLMs' symbolic music reasoning through a benchmark of real-world music scores and user-generated queries, revealing both strengths and challenges.
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