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
Lucen Zhong, Zhengxiao Du, Xiaohan Zhang +2 authors
ComplexFuncBench is a benchmark for evaluating complex function calling in large language models, including multi-step and constrained function calls, with an automatic evaluation framework called ComplexEval.
Ryan Burgert, Yuancheng Xu, Wenqi Xian +10 authors
A new noise warping algorithm improves motion control in video diffusion models by transforming random temporal noise into correlated noise using optical flow, enabling real-time fine-tuning and scalability.
Jiayi Lei, Renrui Zhang, Xiangfei Hu +12 authors
IMAGINE-E evaluates multiple text-to-image models across various domains, highlighting their strengths and limitations, and underscores the expanding applicability of T2I models.
Yantao Liu, Zijun Yao, Rui Min +3 authors
Pairwise Reward Model combined with a knockout tournament enhances test-time scaling of Large Language Models by accurately comparing candidate solutions through parallel evaluation.
Xiaowen Li, Haolan Xue, Peiran Ren +1 authors
DiffuEraser, a video inpainting model based on stable diffusion with enhanced generative and temporal consistency capabilities, outperforms state-of-the-art techniques.
Jianing Yang, Alexander Sax, Kevin J. Liang +6 authors
Fast3R, a Transformer-based extension of DUSt3R, achieves efficient and scalable multi-view 3D reconstruction by processing multiple images in parallel, improving inference speed and accuracy.
Linrui Tian, Siqi Hu, Qi Wang +2 authors
A novel audio-driven talking head method generates expressive facial expressions and hand gestures using a two-stage process, combining direct audio-to-hand-pose generation and a diffusion model for video synthesis.
Chao Feng, Ziyang Chen, Aleksander Holynski +2 authors
GPS conditioning enhances image generation and 3D reconstruction by capturing location-specific details.
Yen-Ting Lin, Di Jin, Tengyu Xu +11 authors
Step-KTO, a training framework providing binary feedback on both intermediate steps and final answers, enhances the reasoning coherence and accuracy of large language models.
Maosong Cao, Taolin Zhang, Mo Li +5 authors
Condor, a two-stage synthetic data generation framework, enhances the performance of large language models through World Knowledge Tree and Self-Reflection Refinement, generating high-quality supervised fine-tuning data.
Nada Saadi, Tathagata Raha, Clément Christophe +3 authors
Training large language models with carefully calibrated language ratios in multilingual medical data improves performance on clinical tasks more effectively than fine-tuning alone.
Di Chang, Hongyi Xu, You Xie +12 authors
X-Dyna uses a diffusion-based pipeline and Dynamics-Adapter to animate human images by integrating facial expressions and body movements, resulting in realistic and expressive videos.
Elad Levi, Ilan Kadar
IntellAgent is a scalable, open-source multi-agent framework that uses policy-driven graph modeling and realistic event generation to comprehensively evaluate conversational AI systems in diverse, synthetic benchmarks.
Shuzhou Yuan, Michael Färber
Hallucinations in Large Language Models can enhance their performance in drug discovery tasks by improving the quality of generated molecular descriptions and task-specific outcomes.
Tao Liu, Kai Wang, Senmao Li +6 authors
A novel training-free method, "One-Prompt-One-Story," enhances consistent identity-preserving text-to-image generation using Singular-Value Reweighting and Identity-Preserving Cross-Attention.
Deyu Zhou, Quan Sun, Yuang Peng +8 authors
MAGI, a hybrid video generation framework using masked and causal modeling, achieves superior performance with Complete Teacher Forcing, overcoming exposure bias and generating long, coherent video sequences.
Paul Röttger, Giuseppe Attanasio, Felix Friedrich +19 authors
MSTS, a Multimodal Safety Test Suite, identifies safety issues in VLMs by evaluating responses to multimodal prompts across various hazard categories.
Shengkui Zhao, Kun Zhou, Zexu Pan +3 authors
HiFi-SR, a unified GAN-based model, achieves high-fidelity speech super-resolution using a transformer-convolutional generator and multi-scale discriminators, outperforming existing methods across various scenarios.
Karthik Viswanathan, Yuri Gardinazzi, Giada Panerai +2 authors
The research explores the relationship between token embedding geometry and next token prediction in transformers, finding correlations between geometric properties and prediction loss.
Giyeong Oh, Saejin Kim, Woohyun Cho +4 authors
SEAL is a watermarking technique for LoRA models that embeds a non-trainable matrix to protect ownership without affecting performance or security.
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