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。
Melisa Russak, Umar Jamil, Christopher Bryant +4 authors
Writing in the Margins (WiM) enhances Large Language Models' performance on long input sequences and retrieval tasks by using chunked prefill and marginal information, improving accuracy and F1-score without fine-tuning.
50 篇论文 · 按点赞排序
Hugo Laurençon, Andrés Marafioti, Victor Sanh +1 authors
A comprehensive tutorial on building vision-language models, detailing the development of Idefics3-8B with an improved dataset enhancing document understanding.
Dani Valevski, Yaniv Leviathan, Moab Arar +1 authors
GameNGen, a neural model-powered game engine, simulates high-quality gameplay in real-time using a diffusion model conditioned on past frames and actions.
Shijia Yang, Bohan Zhai, Quanzeng You +3 authors
Correlation between cross-modal alignment and vision representation improves performance in multimodal large language models, enabling identification and training of optimal vision representation with reduced computational cost.
Min Shi, Fuxiao Liu, Shihao Wang +12 authors
Mixture of vision encoders and resolutions in multimodal large language models improves performance through concatenation of visual tokens and a Pre-Alignment mechanism, leading to superior results on benchmarks.
Trung Dao, Thuan Hoang Nguyen, Thanh Le +4 authors
Enhancements to SwiftBrush through improved weight initialization, efficient LoRA training, and novel clamped CLIP loss achieve state-of-the-art image quality while maintaining generative diversity.
Wenyi Hong, Weihan Wang, Ming Ding +22 authors
The CogVLM2 family of visual language models, including CogVLM2, CogVLM2-Video, and GLM-4V, advances image and video understanding with enhanced architectures and training methods.
Guosheng Dong, Da Pan, Yiding Sun +17 authors
A data processing pipeline and open-sourced details for training a large language model achieve competitive performance with commercial models.
Shengpeng Ji, Ziyue Jiang, Xize Cheng +13 authors
WavTokenizer provides extreme audio compression and state-of-the-art quality by expanding the VQ space, contextual windows, attention networks, and using a multi-scale discriminator and inverse Fourier transform.
Yinghao Ma, Anders Øland, Anton Ragni +40 authors
A review of foundation models in music, including large language models and latent diffusion models, highlights their impact, architectural choices, and the need for ethical considerations in music applications.
Wei Chen, Zhiyuan Li, Shuo Xin +1 authors
Dolphin, a novel decoder-decoder architecture, enables efficient processing of long contexts in language models by distilling contextual information into memory embeddings, thereby improving energy efficiency and latency without sacrificing accuracy.
Junxiong Wang, Daniele Paliotta, Avner May +2 authors
Large-scale Transformer models can be distilled into linear RNNs using attention layer weights, achieving competitive performance with enhanced deployment efficiency.
Daoguang Zan, Zhirong Huang, Ailun Yu +17 authors
A Java version of SWE-bench, SWE-bench-java, has been developed and released to evaluate the issue resolution capabilities of large language models in Java, along with a Docker-based environment and leaderboard.
Zhikai Li, Xuewen Liu, Dongrong Fu +4 authors
K-Sort Arena uses K-wise comparisons and probabilistic modeling to efficiently rank generative models based on human preference, showing significantly faster convergence than the ELO algorithm.
Fangfu Liu, Wenqiang Sun, Hanyang Wang +5 authors
ReconX uses video diffusion models guided by a 3D point cloud to achieve high-quality 3D scene reconstruction from sparse views while maintaining temporal consistency and coherence.
Xiaojuan Wang, Boyang Zhou, Brian Curless +3 authors
A method for generating coherent video sequences between key frames using a dual-directional diffusion process outperforms existing techniques.
Ziyu Guo, Renrui Zhang, Xiangyang Zhu +4 authors
SAM2Point leverages SAM 2 for zero-shot promptable 3D segmentation by interpreting 3D data as multi-directional videos, supporting various prompts and scenarios without additional training or 2D-3D projection.
Tian Ye, Zicheng Xu, Yuanzhi Li +1 authors
Incorporating error-correction data into the pretraining of language models improves reasoning accuracy without multi-round prompting.
Asim Biswal, Liana Patel, Siddarth Jha +5 authors
A unified Table-Augmented Generation (TAG) paradigm addresses limitations of existing Text2SQL and retrieval-based methods, enabling a broader range of natural language queries over databases with improved accuracy and flexibility.
Shuai Yang, Jing Tan, Mengchen Zhang +5 authors
LayerPano3D generates high-quality, full-view consistent, and immersive 3D panoramic scenes from text using layered diffusion priors and Gaussians.
Yi-Fan Zhang, Huanyu Zhang, Haochen Tian +10 authors
MME-RealWorld is a large manually annotated benchmark featuring high-resolution images for evaluating the performance of Multimodal Large Language Models in real-world scenarios.
Chansung Park, Juyong Jiang, Fan Wang +3 authors
The LlamaDuo pipeline enables the transfer of knowledge from cloud-based large language models to smaller, locally manageable models through iterative fine-tuning with synthetic datasets.
Wenhao Li, Yichao Cao, Xie Su +5 authors
ConFiner is an efficient video generation framework that decouples video creation into control and refinement using diffusion models, reducing computational costs while maintaining high quality and coherence.
David Durst, Feng Xie, Vishnu Sarukkai +9 authors
A transformer-based model generates efficient and human-like team movement in CS:GO through data-driven learning from professional gameplay.
Yikang Shen, Matthew Stallone, Mayank Mishra +6 authors
Research reveals a power-law relationship for optimal learning rate, batch size, and training tokens, leading to the development of a Power scheduler that achieves consistent performance across various model sizes and architectures.
Abdelrahman Eldesokey, Peter Wonka
A novel diffusion-based Text-to-Image approach with interactive 3D layout control uses 3D boxes and a Dynamic Self-Attention module to improve object generation and preservation.
Yingyu Liang, Zhizhou Sha, Zhenmei Shi +2 authors
A novel fast computation method for gradient calculation in multi-layer transformer models reduces computational complexity from quadratic to almost linear, improving efficiency and enabling better handling of long-context language models.
Fangxun Shu, Yue Liao, Le Zhuo +13 authors
LLaVA-MoD employs a sparse Mixture of Experts and progressive knowledge transfer to efficiently train small-scale multimodal language models with limited computational resources while outperforming larger models.
Yichao Fu, Siqi Zhu, Runlong Su +3 authors
A novel scheduler using learning to rank improves LLM inference performance by better approximating shortest-job-first scheduling.
Yuncheng Yang, Yulei Qin, Tong Wu +9 authors
An efficient pipeline using few human-annotated samples and a mixture-of-experts system enhances the domain-specific capabilities of large language models with open knowledge.
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