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。
Yujia Qin, Shihao Liang, Yining Ye +15 authors
ToolLLM is a framework that enhances open-source LLMs with tool-use capabilities through a comprehensive dataset, efficient decision tree, automatic evaluation, and neural API retrieval.
42 篇论文 · 按点赞排序
Zhewei Yao, Reza Yazdani Aminabadi, Olatunji Ruwase +16 authors
DeepSpeed-Chat introduces an efficient and accessible RLHF training pipeline for large-scale ChatGPT-like models, enabling rapid and cost-effective training.
SeungHeon Doh, Keunwoo Choi, Jongpil Lee +1 authors
A large-scale pseudo music caption dataset generated using LLMs improves music captioning performance over supervised models in zero-shot and transfer-learning settings.
Xuefei Ning, Zinan Lin, Zixuan Zhou +2 authors
"Skeleton-of-Thought" (SoT) method reduces generation latency and potentially improves answer quality by guiding LLMs to generate a skeleton first, followed by parallel content completion.
Stephen Casper, Xander Davies, Claudia Shi +29 authors
RLHF is analyzed for its limitations and flaws, with suggestions for improvement, complementation, and oversight to enhance safer AI development.
Cheng-Yu Hsieh, Si-An Chen, Chun-Liang Li +5 authors
LLMs can effectively use tool documentation instead of demonstrations to perform complex tasks, achieving performance comparable or superior to few-shot learning on both vision and language benchmarks.
Jessy Lin, Yuqing Du, Olivia Watkins +4 authors
Dynalang, a multimodal agent, learns to predict future text and image representations using language hints to improve task performance and enrich its understanding.
Anas Awadalla, Irena Gao, Josh Gardner +13 authors
OpenFlamingo, an open-source family of autoregressive vision-language models, replicates DeepMind's Flamingo performance across multiple datasets.
Anthony Brohan, Noah Brown, Justice Carbajal +51 authors
Vision-language-action models, trained on large-scale vision and language data, are integrated into robotic control to enhance generalization and semantic reasoning.
Hanjia Lyu, Song Jiang, Hanqing Zeng +2 authors
Combining diverse prompts and input augmentation techniques improves personalized content recommendation performance using large language models.
Michael Moor, Qian Huang, Shirley Wu +6 authors
Med-Flamingo, an adaptation of OpenFlamingo-9B for the medical domain, demonstrates few-shot capabilities in generative visual question answering with significant performance improvements as evaluated by clinicians.
Ning Miao, Yee Whye Teh, Tom Rainforth
A zero-shot verification scheme improves LLMs' question-answering performance by identifying and correcting errors in step-by-step reasoning.
Jiaao Chen, Xiaoman Pan, Dian Yu +4 authors
A novel prompting strategy called skills-in-context (SKiC) enhances compositional generalization in large language models by enabling them to solve unseen complex problems through the synergistic activation and combination of internal skills.
Zheng Yuan, Hongyi Yuan, Chengpeng Li +3 authors
Rejection sampling fine-tuning enhances mathematical reasoning in large language models by generating auxiliary datasets of correct reasoning paths, leading to higher performance and better scalability compared to supervised fine-tuning.
Joan Puigcerver, Carlos Riquelme, Basil Mustafa +1 authors
Soft MoE, a differentiable sparse Transformer, stabilizes training, reduces inference cost, and outperforms traditional Transformers and MoE variants in visual recognition.
Thomas McGrath, Matthew Rahtz, Janos Kramar +2 authors
Analysis reveals adaptive computation and downregulation mechanisms in language models, showing loose coupling among layers even without dropout.
Ke Chen, Yusong Wu, Haohe Liu +3 authors
A text-to-music model named MusicLDM, based on Stable Diffusion and AudioLDM, overcomes limitations in music data and plagiarism with beat tracking and mixup strategies, enhancing music quality, novelty, and text correspondence.
Simon Kornblith, Lala Li, Zirui Wang +1 authors
Classifier-free guidance improves specific caption generation by fine-tuning for both conditional and unconditional distributions, enhancing reference-free metrics and web data quality.
Sarah Schwettmann, Neil Chowdhury, Antonio Torralba
Frozen text transformers can incorporate visual information through self-supervised learning, with specific neurons facilitating translation between modalities and influencing image captioning.
Amir Bar, Florian Bordes, Assaf Shocher +6 authors
FlexPredict enhances self-supervised learning in image processing by incorporating stochastic location uncertainty into masked image modeling, leading to improved performance on tasks like linear probing and video segmentation.
Enxin Song, Wenhao Chai, Guanhong Wang +8 authors
A video understanding system using a dual-memory mechanism based on the Atkinson-Shiffrin model, leveraging Transformers, achieves leading performance on long videos.
Mustafa Shukor, Corentin Dancette, Alexandre Rame +1 authors
UnIVAL, a unified model with 0.25B parameters, efficiently supports text, images, video, and audio, demonstrating competitive performance and benefiting out-of-distribution generalization through multimodal weight interpolation.
Yasheng Sun, Yifan Yang, Houwen Peng +5 authors
ImageBrush uses visual instructions learned through diffusion-based inpainting to perform image manipulation without textual guidance, achieving accurate and generalizable results.
Pierre Fernandez, Antoine Chaffin, Karim Tit +2 authors
This research develops robust watermarking techniques for large language models, introducing statistical tests, comparing watermarks through classical benchmarks, and enhancing detection schemes.
Andrew Guo, Bowen Wen, Jianhe Yuan +4 authors
The HANDAL dataset provides 3D annotations of robotics-ready objects for 6-DoF pose estimation and affordance prediction, enhancing real-world robot interaction scenarios.
Weiyun Wang, Min Shi, Qingyun Li +11 authors
The All-Seeing project introduces a large dataset (AS-1B) and a unified model (ASM) for recognizing and understanding a wide range of concepts in the open world, achieving zero-shot performance across various vision and language tasks.
Shawn Xu, Lin Yang, Christopher Kelly +25 authors
ELIXR, a lightweight adapter architecture combining a language-aligned image encoder with PaLM 2, achieves state-of-the-art performance in zero-shot and data-efficient chest X-ray classification and semantic search, requiring significantly less data compared to existing methods.
Qi Zhao, Ce Zhang, Shijie Wang +4 authors
The proposed AntGPT framework effectively anticipates future actions in videos by integrating large language models to infer goals and predict sequences, achieving state-of-the-art performance on multiple benchmarks.
Junhyeong Cho, Gilhyun Nam, Sungyeon Kim +2 authors
PromptStyler simulates style distribution shifts using prompts in a joint vision-language space for source-free domain generalization, achieving top performance on several benchmarks without images.
Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen +5 authors
UniVTG unifies various video temporal grounding tasks and labels through a unified framework, enabling effective and flexible model development and delivering strong performance across multiple datasets.
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