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
Xuan Ju, Yiming Gao, Zhaoyang Zhang +6 authors
A new video dataset, MiraData, and benchmark, MiraBench, are introduced to improve video generation quality, particularly in motion intensity and temporal consistency, using GPT-4V annotations and DiT-based models.
Chengpeng Li, Guanting Dong, Mingfeng Xue +3 authors
DotaMath, a series of LLMs using decomposition with code assistance and self-correction, achieves high performance on complex mathematical tasks by fine-tuning on a large dataset of interactive tool Use.
Xiaotong Li, Fan Zhang, Haiwen Diao +3 authors
Perceptual Fusion integrates diverse perception experts with an efficient MLLM to generate comprehensive image descriptions, enhancing MLLMs' vision-language capabilities.
Amanda Dsouza, Christopher Glaze, Changho Shin +1 authors
SWiM evaluates long context models and proposes medoid voting to mitigate the lost-in-the-middle effect, significantly improving accuracy in single document QA tasks.
LLM-jp, Akiko Aizawa, Eiji Aramaki +78 authors
LLM-jp, a collaborative project, develops open-source and powerful Japanese large language models with over 1,500 participants.
Jize Wang, Zerun Ma, Yining Li +4 authors
GTA benchmark evaluates LLMs' tool-use capabilities with real user queries, deployed tools, and multimodal inputs, revealing significant challenges and identifying bottlenecks.
Daniel Watson, Saurabh Saxena, Lala Li +2 authors
Cascaded diffusion model for 4D novel view synthesis with joint 3D, 4D, and video training, enabling metric scale camera control and state-of-the-art fidelity with temporal dynamics handling.
Jingwen He, Tianfan Xue, Dongyang Liu +6 authors
VEnhancer enhances video quality by increasing spatial and temporal resolution using a unified video diffusion model and video ControlNet with space-time data augmentation and video-aware conditioning.
Manjie Xu, Chenxing Li, Yong Ren +4 authors
The study explores vision encoders, auxiliary embeddings, and data augmentation to enhance video-to-audio generation quality and synchronization, demonstrating state-of-the-art capabilities with the VTA-LDM model.
Lingwei Meng, Long Zhou, Shujie Liu +9 authors
MELLE, a continuous-valued token-based language model for text-to-speech synthesis, uses regression and variational inference to generate high-fidelity mel-spectrograms directly from text.
Haozhe Zhao, Xiaojian Ma, Liang Chen +7 authors
UltraEdit, a large-scale dataset for instruction-based image editing, improves quality and diversity over existing datasets through real images and region annotations, enhancing diffusion-based editing methods.
Haoji Zhang, Yiqin Wang, Yansong Tang +4 authors
Flash-VStream is a video-language model that efficiently processes online video streams in real-time and responds to user queries, achieving superior performance compared to existing methods on both online and offline video understanding benchmarks.
Xing Han Lù
BM25S provides a fast and efficient Python implementation of BM25 using sparse matrices and eager scoring, achieving significant speedups over popular Python and Java-based frameworks.
Xingyi Yang, Xinchao Wang
Vico is a framework that enhances text-to-video diffusion models by ensuring balanced representation of all concepts in the generated video through spatial-temporal attention graphs and max-flow approximations.
Yutong Wu, Di Huang, Wenxuan Shi +13 authors
A self-generating approach, INVERSE-INSTRUCT, improves instruction-tuned code LLMs by summarizing code to generate additional high-quality instructions, leading to better performance on various benchmarks.
Zhangyang Qi, Yunhan Yang, Mengchen Zhang +7 authors
Tailor3D addresses 3D asset customization by using dual-side images and a novel Dual-sided LRM for seamless integration, enhancing editing efficiency and consistency.
Zachary Kenton, Noah Y. Siegel, János Kramár +8 authors
Debate between AI agents outperforms consultancy in diverse tasks and improves judge accuracy compared to direct question answering when agents choose their answers.
Hila Chefer, Shiran Zada, Roni Paiss +7 authors
A framework for customizing text-to-video models using existing text-to-image customization without video data, employing Spatial Adapters and a Motion Adapter module.
Ruida Wang, Jipeng Zhang, Yizhen Jia +4 authors
TheoremLlama is an end-to-end framework using Large Language Models to generate formal Lean4 proofs, demonstrating higher accuracy than GPT-4 on theorem proving tasks with an NL-FL aligned dataset.
Frederic Z. Zhang, Paul Albert, Cristian Rodriguez-Opazo +2 authors
aTLAS algorithm leverages parameter blocks from pre-trained models to enhance knowledge composition and transfer, achieving better generalization and flexibility with limited data.
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