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
Yujie Wei, Shiwei Zhang, Hangjie Yuan +9 authors
DreamVideo-2 generates customized videos using a single image and bounding box sequence without test-time fine-tuning, balancing subject learning and motion control through masked reference attention and reweighted diffusion loss.
Peng Xia, Kangyu Zhu, Haoran Li +6 authors
A versatile multimodal RAG system, MMed-RAG, enhances the factuality of Medical Large Vision-Language Models across different medical domains by introducing domain-aware retrieval, adaptive context selection, and preference fine-tuning.
Chenxi Wang, Xiang Chen, Ningyu Zhang +4 authors
A novel dynamic correction decoding method for multimodal large language models reduces hallucination rates by integrating knowledge from preceding layers into the final output.
Dejia Xu, Yifan Jiang, Chen Huang +5 authors
Cavia is a novel framework that integrates view-integrated attention modules to generate camera-controllable, multi-view, spatiotemporally consistent videos from images.
Zichen Zhu, Hao Tang, Yansi Li +8 authors
MobA, a mobile assistant using multimodal large language models, improves task execution through a dual-agent architecture with reflection, handling complex instructions effectively.
Mingchen Zhuge, Changsheng Zhao, Dylan Ashley +10 authors
Agent-as-a-Judge framework uses agentic systems for evaluation, offering intermediate feedback and reliability improvements over existing methods like LLM-as-a-Judge in tasks such as code generation.
Tianyi Bai, Ling Yang, Zhen Hao Wong +8 authors
A multi-agent collaborative mechanism enhances data selection for pretraining large language models, improving efficiency and performance.
Peng Jin, Bo Zhu, Li Yuan +1 authors
Mixture-of-Head attention enhances Transformer models by efficiently selecting relevant attention heads, improving performance with fewer resources.
Hongyu Zhao, Ming Li, Lichao Sun +1 authors
Efficiently reducing the number of tasks for evaluating large language models with minimal impact on evaluation quality using a transferability metric based on in-context learning.
Yedi Hu, Yunzhi Yao, Ningyu Zhang +2 authors
Model kinship guides model merging to improve performance on benchmarks and helps escape local optima in model evolution.
Tianhao Wu, Janice Lan, Weizhe Yuan +3 authors
An iterative search and optimization procedure equips LLMs with thinking abilities for general instruction following, improving performance across various categories, including reasoning, marketing, health, and general knowledge.
Siwei Wu, Zhongyuan Peng, Xinrun Du +14 authors
OpenAI's o1 model enhances LLM reasoning capabilities through Test-time Compute methods, outperforming existing strategies across math, coding, and commonsense reasoning tasks.
Zekun Moore Wang, Shawn Wang, Kang Zhu +5 authors
PopAlign enhances large language model alignment by introducing diverse contrasting patterns across prompts, models, and pipelines, leading to improved alignment compared to traditional methods.
Pei Wang, Yanan Wu, Zekun Wang +12 authors
A multi-granularity tool-use benchmark, MTU-Bench, is proposed to evaluate large language models across various tool usage scenarios, costing effectively without relying on GPT or human metrics.
Rui Zhao, Hangjie Yuan, Yujie Wei +8 authors
EvolveDirector leverages pre-trained vision-language models to guide and refine a text-to-image generation model, reducing data requirements and improving performance compared to advanced models.
Yahan Li, Tingyu Xia, Yi Chang +1 authors
The Matrix Nuclear-Norm is introduced as an efficient and scalable metric for evaluating large language models by approximating matrix rank, reducing computational complexity and improving performance assessment.
Tianwei Xiong, Yuqing Wang, Daquan Zhou +3 authors
A new dataset, LVD-2M, is curated to enhance long video generation by providing high-quality, long-take videos with temporally dense captions.
Yuxuan Sun, Ruikang Liu, Haoli Bai +10 authors
FlatQuant, a novel post-training quantization method, achieves better accuracy and faster inference compared to existing methods by finding optimal affine transformations and leveraging Kronecker decomposition.
Cheng Lu, Yang Song
A new theoretical framework and algorithm improve continuous-time consistency models, enabling large-scale training and achieving near-state-of-the-art FID scores.
Zhiyuan Ma, Yuzhu Zhang, Guoli Jia +7 authors
This survey provides an efficiency-focused overview of diffusion model principles, architecture design, training, inference, and deployment to advance research and application.
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