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
Xiyao Wang, Zhengyuan Yang, Chao Feng +6 authors
A method using Monte Carlo Tree Search for selecting challenging training samples enhances visual reasoning capabilities in a smaller dataset compared to existing models.
Yiting Lu, Jiakang Yuan, Zhen Li +15 authors
OmniCaptioner generates detailed captions across various visual domains, enhancing visual reasoning with LLMs, improving image generation tasks, and enabling efficient supervised fine-tuning.
Andreas Hochlehnert, Hardik Bhatnagar, Vishaal Udandarao +3 authors
A standardized evaluation framework reveals that reinforcement learning offers limited improvements over supervised fine-tuning in mathematical reasoning benchmarks, highlighting issues with transparency and variability in current methodologies.
Shuzhang Zhong, Yanfan Sun, Ling Liang +3 authors
HybriMoE enhances the hybrid CPU-GPU inference of MoE models through dynamic scheduling, prefetching, and caching to improve resource utilization and reduce expert activation instability.
Xianwei Zhuang, Yuxin Xie, Yufan Deng +5 authors
VARGPT-v1.1, an advanced visual autoregressive model, integrates iterative visual instruction tuning with reinforcement learning, enhances resolution, and achieves state-of-the-art performance in multimodal understanding and text-to-image tasks.
Gabriel Grand, Joshua B. Tenenbaum, Vikash K. Mansinghka +2 authors
DisCIPL, a method combining a Planner and Follower model, enables efficient and verifiable reasoning in language models by generating task-specific inference programs.
Yixing Li, Ruobing Xie, Zhen Yang +8 authors
TransMamba combines Transformers and Mamba frameworks to leverage the strengths of both, achieving efficient and effective long-sequence processing through shared parameters and dynamic switching mechanisms.
Ming Li, Ruiyi Zhang, Jian Chen +6 authors
A new task, TRIG, with an instruction dataset benchmarking MLLMs on text-rich document images improves grounding capabilities through OCR-LLM-human interaction and finetuning.
Genglin Liu, Salman Rahman, Elisa Kreiss +2 authors
A social network simulation framework using generative language agents models user behavior, misinformation mitigation, and content engagement, with open-source software provided.
Akshara Prabhakar, Zuxin Liu, Weiran Yao +12 authors
The APIGen-MT framework generates high-quality multi-turn agent data using detailed task blueprints and simulated human-agent interactions, leading to superior model performance and consistency.
Runnan Fang, Xiaobin Wang, Yuan Liang +8 authors
SynWorld framework enables agents to autonomously explore environments and refine action knowledge through scenario synthesis and Monte Carlo Tree Search.
Zhaorun Chen, Mintong Kang, Bo Li
ShieldAgent enforces safety policy compliance for autonomous agents using logical reasoning and probabilistic rule circuits, outperforming existing methods on benchmarks while reducing computational overhead.
Yunlong Tang, Jing Bi, Chao Huang +16 authors
CAT-V provides detailed video object-centric captions using spatiotemporal prompts without additional training, integrating segmentation, temporal analysis, and caption generation.
Jing Bi, Susan Liang, Xiaofei Zhou +15 authors
This paper provides an overview of reasoning techniques in large language models (LLMs) handling both textual and multimodal inputs, detailing challenges, opportunities, and optimization strategies.
Mingyang Fu, Yuyang Peng, Benlin Liu +2 authors
Evaluation of various MLLMs on LiveVQA, a dataset of visual questions with up-to-date visual knowledge, shows that advanced visual reasoning is essential for complex multi-hop questions.
Jinqi Luo, Tianjiao Ding, Kwan Ho Ryan Chan +3 authors
Concept Lancet (CoLan) is a framework for diffusion-based image editing that improves performance and consistency by decomposing source images into latent concept representations.
Boyuan Wang, Runqi Ouyang, Xiaofeng Wang +6 authors
A novel framework, HumanDreamer-X, enhances single-image human reconstruction through unified 3D Gaussian Splatting and HumanFixer, improving PSNR and handling in-the-wild data.
Wulin Xie, Yi-Fan Zhang, Chaoyou Fu +6 authors
A new evaluation framework assesses Unified MLLMs across traditional and mixed-modality tasks, revealing performance gaps and the need for more robust models.
Xinhao Li, Ziang Yan, Desen Meng +7 authors
Reinforcement Fine-Tuning with Group Relative Policy Optimization enhances video multimodal large language models, improving spatio-temporal perception without compromising chat abilities.
Jiazi Bu, Pengyang Ling, Yujie Zhou +7 authors
HiFlow is a training-free framework that improves high-resolution image synthesis of text-to-image models by aligning pre-trained flow models across various aspects.
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