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。
Guowei Xu, Peng Jin, Ziang Wu +5 authors
LLaVA-CoT is a vision-language model that achieves improved reasoning performance through structured multistage processing and test-time scaling, outperforming larger models with a smaller training dataset.
50 篇论文 · 按点赞排序
Siming Huang, Tianhao Cheng, Jason Klein Liu +16 authors
OpenCoder is a top-tier open-access code LLM that provides comprehensive reproducible resources, including data, pipelines, and protocols, to advance code AI research.
Kevin Qinghong Lin, Linjie Li, Difei Gao +6 authors
ShowUI is a vision-language-action model that enhances GUI assistants by using UI-guided token selection and interleaved vision-language-action streaming, achieving high accuracy and efficiency in zero-shot screenshot grounding across different environments.
Yuchao Gu, Yipin Zhou, Yunfan Ye +5 authors
ROICtrl, an adapter for diffusion models using ROI-Align and ROI-Unpool, improves regional instance control in text-based visual generation with lower computational costs.
Weiyun Wang, Zhe Chen, Wenhai Wang +8 authors
A preference optimization process enhances multimodal reasoning, particularly Chain-of-Thought performance, in large language models through automated data construction and mixed preference optimization, demonstrating significant improvements on benchmarks.
Viacheslav Surkov, Chris Wendler, Mikhail Terekhov +3 authors
Sparse autoencoders are applied to learn interpretable features from intermediate updates in text-to-image diffusion models, such as SDXL Turbo, enhancing understanding of these models' internal mechanisms.
Zichen Liu, Yue Yu, Hao Ouyang +6 authors
MagicQuill integrates an interface with a multimodal large language model and diffusion prior to enable efficient and precise real-time image editing through minimal user input.
Zhengyi Wang, Jonathan Lorraine, Yikai Wang +4 authors
The work demonstrates the capability of LLMs to generate 3D meshes from text by introducing a novel approach to tokenize 3D mesh data, allowing the unification of 3D and text modalities without expanding the model's vocabulary.
Taiming Lu, Tianmin Shu, Alan Yuille +2 authors
Generative World Explorer (Genex) enables agents to mentally explore 3D environments using imagined observations, updating their beliefs without physical exploration to improve decision-making.
Jiejun Tan, Zhicheng Dou, Wen Wang +3 authors
HtmlRAG enhances Retrieval-Augmented Generation (RAG) systems by using HTML instead of plain text, improving knowledge modeling and reducing information loss through HTML cleaning, compression, and pruning.
David Junhao Zhang, Roni Paiss, Shiran Zada +7 authors
ReCapture generates new videos with different camera trajectories from a user-provided video using multiview diffusion models and masked video fine-tuning.
Hongyu Wang, Shuming Ma, Furu Wei
BitNet a4.8 enhances the efficiency of large language models through 4-bit quantization and sparsification, achieving equivalent performance to BitNet b1.58 with reduced inference costs.
Antoine Grosnit, Alexandre Maraval, James Doran +15 authors
Agent K v1.0 automates data science tasks across various domains using structured reasoning and Bayesian optimization, achieving performance comparable to Expert-level Kaggle users.
Nathan Lambert, Jacob Morrison, Valentina Pyatkin +20 authors
T\"ULU 3, an open-source family of post-trained language models, introduces transparent training data, recipes, and advanced techniques to match or surpass proprietary models in performance.
Yoad Tewel, Rinon Gal, Dvir Samuel Yuval Atzmon +2 authors
Add-it extends diffusion models with a weighted extended-attention mechanism to seamlessly integrate objects into images based on text instructions, achieving state-of-the-art results on real and generated image insertion benchmarks.
Siheng Li, Cheng Yang, Zesen Cheng +4 authors
A novel method enables LLMs to self-improve in long-context reasoning through sampling, scoring with Minimum Bayes Risk, and supervised fine-tuning, outperforming expert-annotated data approaches.
Ming Li, Yanhong Li, Tianyi Zhou
Investigating gradient patterns in LLMs during fast vs. slow training reveals that slow thinking (chain-of-thoughts) enhances learning stability and distinguishes correct reasoning paths.
Yu Zhao, Huifeng Yin, Bo Zeng +6 authors
Marco-o1 extends reasoning models to open-ended domains without clear standards using Chain-of-Thought fine-tuning, Monte Carlo Tree Search, reflection mechanisms, and innovative reasoning strategies.
Zhenxiong Tan, Songhua Liu, Xingyi Yang +2 authors
OminiControl, a parameter-efficient framework, integrates image conditions into Diffusion Transformer models using a parameter reuse mechanism, excelling in subject-driven and spatially-aligned conditional generation.
Rundi Wu, Ruiqi Gao, Ben Poole +4 authors
CAT4D is a method that uses a multi-view video diffusion model to generate 4D dynamic scenes from monocular video by optimizing a deformable 3D Gaussian representation.
Maurice Weber, Daniel Fu, Quentin Anthony +16 authors
The RedPajama datasets are introduced to address core challenges for open-source language models by providing transparent data curation, large volumes of high-quality text, and quality signals for web data analysis.
Jintao Zhang, Haofeng Huang, Pengle Zhang +3 authors
SageAttention2 improves attention computation speed with 4-bit matrix multiplication and adaptive quantization, achieving significant OPS increases without compromising accuracy.
Wenqiang Sun, Shuo Chen, Fangfu Liu +4 authors
DimensionX leverages video diffusion with spatial-temporal decoupling and trajectory-aware mechanisms to generate high-quality 3D and 4D scenes from single images.
Shantanu Acharya, Fei Jia, Boris Ginsburg
Star Attention improves inference efficiency of large language models on long sequences by using block-sparse approximation, reducing memory and time without significant accuracy loss.
Eldar Kurtic, Alexandre Marques, Shubhra Pandit +2 authors
A comprehensive study evaluates quantization formats for LLMs, showing that FP8 W8A8-FP is lossless, W8A8-INT incurs minimal accuracy loss, and W4A16-INT is cost-efficient, providing guidelines for optimal deployment.
Xin Huang, Tengfei Wang, Ziwei Liu +1 authors
Material Anything is a unified diffusion framework that generates physically-based materials for 3D objects using a pre-trained image diffusion model with enhancements for stability and quality.
Weixin Liang, Lili Yu, Liang Luo +8 authors
Mixture-of-Transformers (MoT) reduces pretraining computational costs for multi-modal large language models by decoupling non-embedding parameters by modality, enabling efficient processing and competitive performance across text, image, and speech.
Xin Dong, Yonggan Fu, Shizhe Diao +10 authors
Hymba, a family of small language models with a hybrid-head architecture combining transformer attention and state space models, achieves state-of-the-art performance with improved efficiency and reduced cache size.
Cong Wei, Zheyang Xiong, Weiming Ren +3 authors
An omnipotent image editor trained on multisource supervision, importance sampling, and a versatile architecture excels in multiple tasks with diverse aspect ratios and high data quality.
Dingjie Song, Sicheng Lai, Shunian Chen +2 authors
A framework called MM-Detect is introduced to detect data contamination in multimodal large language models, demonstrating sensitivity across different contamination levels and identifying potential contamination stages during pre-training and fine-tuning.
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