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。
50 篇论文 · 按点赞排序
Pier Giuseppe Sessa, Robert Dadashi, Léonard Hussenot +17 authors
BOND, a novel RLHF algorithm, distills the Best-of-N sampling strategy without significant computational overhead, enhancing the quality of generative language models.
Ofir Abramovich, Niv Nayman, Sharon Fogel +7 authors
VisFocus enhances visual document understanding by integrating language prompts directly into the vision encoder, enabling focus on relevant text patches and improving performance across benchmarks.
Zuyan Liu, Benlin Liu, Jiahui Wang +5 authors
Elastic Cache is introduced to enhance the efficiency of instruction-following large vision-language models by applying importance-driven cache merging strategies in specific stages of the model.
Bo Liu, Rui Wang, Lemeng Wu +3 authors
A novel deep state-space model based on implicit update rules outperforms existing models in sequence and language modeling tasks.
Yiming Xie, Chun-Han Yao, Vikram Voleti +2 authors
Stable Video 4D generates temporally consistent novel views for dynamic 3D objects using a unified latent video diffusion model, outperforming prior methods in both novel-view video synthesis and 4D generation.
Chih-Hai Su, Chih-Yao Hu, Shr-Ruei Tsai +3 authors
BoostMVSNeRFs enhances MVS-based NeRFs by improving rendering quality in large-scale scenes through advanced cost volume selection and combination techniques.
Minyang Tian, Luyu Gao, Shizhuo Dylan Zhang +27 authors
SciCode, a scientist-curated coding benchmark, evaluates language models' ability to solve scientific research problems across various fields, highlighting current capabilities and future challenges in AI-assisted science.
Zhecheng Yuan, Tianming Wei, Shuiqi Cheng +3 authors
Maniwhere, a visual reinforcement learning framework, enables robotic policies to generalize across varied visual disturbances using multi-view representation and Spatial Transformer Network, demonstrating superior performance in sim2real transfer.
Jie Yang, Xuesong Niu, Nan Jiang +2 authors
Semantic-HOI dataset and F-HOI model improve 3D human-object interaction representation by focusing on detailed state-level descriptions and multimodal instructions.
Itai Gat, Tal Remez, Neta Shaul +5 authors
Discrete Flow Matching is a novel generative model for discrete data that achieves high performance without an autoregressive approach, improving benchmarks like HumanEval and MBPP.
Ruixiang Jiang, Changwen Chen
Artist is a training-free method to control a pretrained diffusion model for text-driven stylization, achieving aesthetic results by separating and sharing information between content and style denoising processes.
Zhiyang Zhang, Fangkai Yang, Xiaoting Qin +6 authors
An LLM-based multi-agent framework enhances microservice management towards achieving autonomic computing, demonstrating progress in Autonomous Service Maintenance.
Yewon Lim, Changyeon Lee, Aerin Kim +1 authors
Knowledge distillation from diffusion models is used to develop a more efficient and lightweight deepfake detector compared to the DIRE framework.
Sogand Salehi, Mahdi Shafiei, Teresa Yeo +2 authors
A method to personalize image generation using user feedback and a large language model to guide a text-to-image model.
Paul Balança, Sam Hosegood, Carlo Luschi +1 authors
Scalify is an end-to-end scale propagation paradigm that supports float8 matrix multiplication and float16 optimizer state storage, improving low-precision training in large language models.
Adrian Remonda, Nicklas Hansen, Ayoub Raji +4 authors
A simulation platform for autonomous racing using Assetto Corsa tests reinforcement learning and MPC algorithms, providing a realistic benchmarking environment.
Yiwei Ma, Zhibin Wang, Xiaoshuai Sun +4 authors
INF-LLaVA, a novel Multimodal Large Language Model, uses a Dual-perspective Cropping Module and Dual-perspective Enhancement Module to effectively process high-resolution images, capturing both local details and global context.
Shayne Longpre, Robert Mahari, Ariel Lee +46 authors
The audit of web domains used for training AI systems reveals a growing inconsistency and restriction of data use, posing a significant threat to the diversity and availability of open web data for AI.
Haiyang Zhou, Xinhua Cheng, Wangbo Yu +2 authors
HoloDreamer uses a stylized panorama generation pipeline and 3D Gaussian Splatting to create view-consistent and fully enclosed 3D scenes from text prompts, outperforming existing methods in visual consistency and reconstruction quality.
Han Guo, William Brandon, Radostin Cholakov +3 authors
FLUTE, a lookup table engine for quantized language models, accelerates inference by minimizing bit manipulations and optimizing shared memory usage, achieving faster performance compared to existing kernels.
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