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。
43 篇论文 · 按点赞排序
Zhenzhen Chu, Jiayu Chen, Cen Chen +4 authors
DualToken-ViT combines the strengths of CNNs and ViTs to achieve efficient vision processing with enhanced global and position-aware features.
Guy Yariv, Itai Gat, Sagie Benaim +3 authors
A lightweight adaptor network improves video generation conditioned on audio and text by aligning audio segments with video segments and enhancing alignment and diversity.
Nikita Janakarajan, Tim Erdmann, Sarath Swaminathan +2 authors
Language models, particularly transformers, are used in molecular discovery, enhancing drug design, property prediction, and reaction chemistry with open-source tools, and envisioning future chatbot interfaces to computational chemistry.
Qiao Gu, Alihusein Kuwajerwala, Sacha Morin +13 authors
ConceptGraphs, a graph-structured representation leveraging 2D foundation models, enables semantically rich and scalable 3D scene understanding for task-driven perception and planning.
Ruoyu Feng, Wenming Weng, Yanhui Wang +5 authors
CCEdit is a framework that decouples video structure and appearance, utilizes ControlNet for structure preservation, and employs adaptable temporal modules and personalization techniques to enable precise and creative video editing.
Chethan Bhateja, Derek Guo, Dibya Ghosh +6 authors
A system called V-PTR integrates temporal-difference learning from video data to enhance robotic offline reinforcement learning, leading to better policies for manipulation tasks.
Jiahao Xie, Wei Li, Xiangtai Li +3 authors
MosaicFusion improves instance segmentation by generating synthetic labeled data through a diffusion-based approach using text-to-image models and cross-attention maps.
Jiamin Xie, Ke Li, Jinxi Guo +7 authors
An adaptive masking approach efficiently prunes multilingual ASR models, outperforming existing methods and reducing language-specific pruning need.
Emanuele Aiello, Lili Yu, Yixin Nie +2 authors
The Joint Autoregressive Mixture (JAM) framework integrates text and image generation models for high-quality multimodal output, exceeding previous methods with a specialized instruction-tuning strategy.
Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy +3 authors
A novel 3D contact detector, DECO, infers dense, vertex-level contact between the full body surface and objects in arbitrary images using body-part-driven and scene-context-driven attention, outperforming existing methods across benchmarks.
Ida Momennejad, Hosein Hasanbeig, Felipe Vieira +5 authors
Systematic evaluation of cognitive maps and planning abilities in multiple large language models reveals significant failures in planning tasks and lack of understanding of underlying cognitive structures.
Mert Yuksekgonul, Varun Chandrasekaran, Erik Jones +5 authors
Transformer-based LLMs show a positive correlation between attention to constraint tokens and factual accuracy, and SAT Probe can predict factual errors by analyzing self-attention patterns.
Jiachen Sun, Mark Ibrahim, Melissa Hall +4 authors
Visual Prompt Adaptation (VPA) enhances the performance of natural language processing models on various tasks through test-time adaptation using learnable tokens, improving OOD generalization, corruption robustness, and domain adaptation.
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