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。
Timothée Darcet, Maxime Oquab, Julien Mairal +1 authors
Additional input tokens in Vision Transformers mitigate artifacts in feature maps, enhancing performance and enabling smoother visual processing.
43 篇论文 · 按点赞排序
Ramakrishna Bairi, Atharv Sonwane, Aditya Kanade +6 authors
CodePlan automates repository-level coding tasks, such as package migration and temporal code edits, using a planning framework that leverages LLMs with context derived from code repositories and change analysis.
Seungwhan Moon, Andrea Madotto, Zhaojiang Lin +10 authors
AnyMAL is a unified model that processes multiple modalities (text, image, video, audio, IMU) and generates text, achieving top performance in multimodal tasks through a pre-trained aligner and fine-tuning with diverse instructions.
Jiaxiang Tang, Jiawei Ren, Hang Zhou +2 authors
DreamGaussian is a 3D content generation framework that combines generative Gaussian splatting with mesh extraction and texture refinement for high efficiency and quality.
Yuhui Xu, Lingxi Xie, Xiaotao Gu +6 authors
QA-LoRA algorithm reduces computational requirements for large language models by quantizing weights during fine-tuning and integrating auxiliary weights post-training.
Yaohui Wang, Xinyuan Chen, Xin Ma +17 authors
LaVie, a text-to-video framework integrating cascaded video latent diffusion models, achieves state-of-the-art video generation by leveraging temporal self-attentions and joint image-video fine-tuning.
Jinze Bai, Shuai Bai, Yunfei Chu +45 authors
Qwen, a series of large language models, including chat and specialized coding and mathematics variants, exhibit superior performance across various tasks and outperform open-source models.
Kumar Tanmay, Aditi Khandelwal, Utkarsh Agarwal +1 authors
A new evaluation framework connects human psychology and AI to assess the moral reasoning and development stages of large language models using a Psychometric Assessment Tool.
Han Lin, Abhay Zala, Jaemin Cho +1 authors
VideoDirectorGPT combines LLMs for video content planning with a layout-controlled video generator, enabling temporally consistent multi-scene video generation from text prompts.
Zilong Chen, Feng Wang, Huaping Liu
Gaussian Splatting based text-to-3D generation improves geometry accuracy and fidelity through progressive optimization, including geometry optimization and appearance refinement.
Xiaoliang Dai, Ji Hou, Chih-Yao Ma +23 authors
Quality-tuning enhances the aesthetic appeal of text-to-image models through supervised fine-tuning with a small set of high-quality images.
Zhiqing Sun, Sheng Shen, Shengcao Cao +9 authors
The paper presents Factually Augmented RLHF to address multimodal misalignment in large multimodal models, significantly improving hallucination reduction and overall performance on vision-language tasks compared to existing methods.
Wenhan Xiong, Jingyu Liu, Igor Molybog +18 authors
Long-context LLMs achieve significant advancements in handling extended contexts through continual pretraining and efficient instruction tuning, surpassing existing models on various benchmarks.
Li Siyao, Tianpei Gu, Weiye Xiao +3 authors
AnimeInbet addresses line inbetweening by geometrically representing line drawings as graphs and using a vertex correspondence Transformer and visibility predictor for high-quality frame interpolation.
Fabian Mentzer, David Minnen, Eirikur Agustsson +1 authors
The proposed finite scalar quantization (FSQ) for VAE latent representations offers competitive performance in image generation and other vision tasks without the complexity of VQ.
Yu Yu, Chao-Han Huck Yang, Jari Kolehmainen +15 authors
Rescore-BERT, using low-rank adaptation, achieves efficient domain-specific speech recognition output rescoring with reduced training parameters and time.
Zhewei Yao, Xiaoxia Wu, Conglong Li +6 authors
A framework called DeepSpeed-VisualChat optimizes LLMs for multi-modal, multi-image, and multi-round dialogues using an innovative causal attention mechanism and data blending techniques.
Sam Ade Jacobs, Masahiro Tanaka, Chengming Zhang +4 authors
DeepSpeed-Ulysses enables efficient and scalable training of Transformer-based LLMs with long sequences by partitioning input data along the sequence dimension and using all-to-all communication.
Mitchell Wortsman, Peter J. Liu, Lechao Xiao +13 authors
Training instabilities in large Transformer models are reproducible in smaller models at high learning rates, with similar mitigations effective; scaling behaviors of activations and gradients predict instabilities.
Hu Xu, Saining Xie, Xiaoqing Ellen Tan +7 authors
MetaCLIP, a metadata-driven data curation approach for language-image pre-training, outperforms CLIP and scales effectively with larger datasets.
David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu +5 authors
Show-1, a hybrid model combining pixel-based and latent-based VDMs, generates high-quality text-to-video with precise alignment and reduced computational cost.
Jan Hendrik Metzen, Piyapat Saranrittichai, Chaithanya Kumar Mummadi
AutoCLIP auto-tunes zero-shot classifiers by assigning per-image weights to prompt templates, improving performance across vision-language models and datasets.
Kumar Shridhar, Harsh Jhamtani, Hao Fang +3 authors
A modular framework SCREWS for reasoning with revisions in large language models reveals novel strategies and improves performance across various reasoning tasks by enabling selection and heterogeneous approaches.
Ari Seff, Brian Cera, Dian Chen +6 authors
MotionLM, a language modeling approach, forecasts multi-agent motion by generating joint distributions over interactive agent futures, achieving state-of-the-art performance on the Waymo Open Motion Dataset.
Luming Tang, Nataniel Ruiz, Qinghao Chu +8 authors
RealFill is a personalized generative inpainting model that fills in missing parts of images with content faithful to the original scene using a few reference images.
Zhang Chen, Zhong Li, Liangchen Song +4 authors
A new type of neural fields using general radial bases with adaptive positions and shapes outperforms existing methods in signal representation, spatial adaptivity, and rendering quality.
Shen Zheng, Yuyu Zhang, Yijie Zhu +4 authors
GPT-Fathom evaluates leading LLMs using a standardized suite to provide insights into the evolution from GPT-3 to GPT-4, addressing aspects like reasoning capabilities, SFT, RLHF, and alignment.
Yuxuan Liu, Tianchi Yang, Shaohan Huang +6 authors
AutoCalibrate uses an iterative, gradient-free approach to automatically align an off-the-shelf LLM-based evaluator with human preferences, improving its correlation with expert evaluations.
Haihao Shen, Naveen Mellempudi, Xin He +3 authors
Research demonstrates that FP8 data formats improve quantization in deep learning models, outperforming INT8 in accuracy and suitability across various tasks.
Antoine Yang, Arsha Nagrani, Ivan Laptev +2 authors
The introduction of VidChapters-7M, a large-scale dataset for video chaptering, improves state-of-the-art performance on video-language tasks and demonstrates the benefit of pretraining on extensive 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