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。
42 篇论文 · 按点赞排序
Xiaoxuan Wang, Ziniu Hu, Pan Lu +7 authors
SciBench introduces an extensive benchmark suite to evaluate the reasoning capabilities of large language models on complex scientific problems, revealing areas for improvement in these models' performance.
Stefan Lionar, Xiangyu Xu, Min Lin +1 authors
A new method, NU-MCC, addresses limitations in 3D reconstruction by using a Neighborhood decoder and a Repulsive Unsigned Distance Function, improving inference speed and reconstruction quality compared to the current state-of-the-art method.
Samyadeep Basu, Maziar Sanjabi, Daniela Massiceti +2 authors
A method called SDS-CLIP improves compositional visio-linguistic reasoning in CLIP using distillation from large text-to-image generative models like Stable-Diffusion.
Borja Rodríguez-Gálvez, Arno Blaas, Pau Rodríguez +5 authors
A new entropy and reconstruction (ER) bound is introduced to analyze multi-view self-supervised learning methods, showing its relation to clustering and distillation approaches, and demonstrating competitive performance with enhanced stability.
Chen Sun, Calvin Luo, Xingyi Zhou +2 authors
End-to-end visual reasoning is achieved using a self-supervised transformer-based framework that compresses video frames into tokens and reconstructs them, outperforming supervised pretraining methods.
Wei Cheng, Ruixiang Chen, Wanqi Yin +18 authors
DNA-Rendering is a comprehensive dataset of human performance data that enhances neural actor rendering accuracy by providing diverse assets and a professional multi-view capture system.
Nilesh Kulkarni, Davis Rempe, Kyle Genova +4 authors
A framework uses neural interaction fields and diffusion models to generate realistic 3D human motions interacting with objects using limited data.
Weifeng Lin, Ziheng Wu, Jiayu Chen +2 authors
A Scale-Aware Modulation Transformer combines convolutional networks with attention mechanisms using Multi-Head Mixed Convolution and Scale-Aware Aggregation modules, achieving superior performance across visual tasks compared to existing models.
Shuhan Tan, Boris Ivanovic, Xinshuo Weng +2 authors
LCTGen, a combination of a large language model and transformer-based decoder, generates realistic dynamic traffic scenes by selecting map locations and modeling vehicle dynamics.
Shiyang Li, Jun Yan, Hai Wang +4 authors
The study proposes verbalizer manipulation to evaluate models' ability to follow instructions by using verbalizers with varying alignment to the model's priors, revealing significant differences in instruction-following capability among models.
Ettore Randazzo, Alexander Mordvintsev
Biomaker CA utilizes cellular automata for simulating plant growth and evolution in virtual biomes, allowing for complex environmental modeling and user interaction.
Pranshu Malviya, Gonçalo Mordido, Aristide Baratin +5 authors
A memory-augmented version of Adam optimizer enhances generalization by encouraging exploration towards flatter minima in the loss landscape.
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