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
Haoyu Liu, Shaohan Huang, Jianfeng Liu +6 authors
A new retrieval method, GeAR, enhances document retrieval by focusing on fine-grained semantic relationships using fusion and decoding modules, achieving competitive performance with minimal additional computational cost.
Aobo Kong, Wentao Ma, Shiwan Zhao +7 authors
Segment-Level Direct Preference Optimization (SDPO) enhances complex goal-oriented social dialogues in LLM-powered social agents by focusing on key segments, outperforming existing DPO methods.
Kam Woh Ng, Jing Yang, Jia Wei Sii +5 authors
A system using multi-view diffusion and continuous part distributions generates novel, plausible 3D objects with fine-grained details that go beyond current examples.
Jiazheng Xu, Yu Huang, Jiale Cheng +18 authors
A reward model with multi-dimensional analysis aligns visual generation models with human preferences, surpassing existing methods in image and video scoring.
Gregor Geigle, Florian Schneider, Carolin Holtermann +4 authors
A comprehensive investigation into training strategies for multilingual vision-language models reveals optimal ways to include multiple languages without degrading English performance and introduces a new benchmark for text-in-image understanding.
Yaoxiang Wang, Haoling Li, Xin Zhang +10 authors
A feature tree-based synthesis framework inspired by AST enhances code LLMs by generating diverse and nuanced code, resulting in state-of-the-art performance across various benchmarks.
Xiaohui Chen, Yinkai Wang, Jiaxing He +4 authors
A Graph Generative Pre-trained Transformer (G2PT) using node and edge sequences outperforms existing graph generative models and demonstrates versatility in tasks like molecular design and property prediction.
Zixuan Huang, Mark Boss, Aaryaman Vasishta +2 authors
SPAR3D combines efficient point cloud generation with detailed mesh creation for high-fidelity single-image 3D reconstruction.
Yanjiang Liu, Shuhen Zhou, Yaojie Lu +6 authors
Auto-RT, a reinforcement learning framework, enhances the detection of complex vulnerabilities in large language models by optimizing attack strategies through early exploration and progressive reward tracking.
Run Luo, Ting-En Lin, Haonan Zhang +10 authors
Openomni, a two-stage training method combining omnimodal alignment and speech generation, achieves state-of-the-art performance in generating natural, emotion-rich dialogues and real-time emotional speech.
Jiakang Yuan, Xiangchao Yan, Botian Shi +6 authors
Dolphin is an auto-research framework that generates and tests scientific research ideas, comparing to state-of-the-art methods in image and point classification tasks.
Ziyang Song, Zerong Wang, Bo Li +5 authors
DepthMaster, a single-step diffusion model with Feature Alignment and Fourier Enhancement modules, achieves state-of-the-art performance in monocular depth estimation by balancing generative and discriminative features.
Weikang Bian, Zhaoyang Huang, Xiaoyu Shi +3 authors
A novel framework combining pseudo 4D Gaussian fields with a video Diffusion Transformer (DiT) allows for more efficient and flexible 4D video generation, supporting advanced cinematic effects and different camera parameters.
Yuechen Zhang, Yaoyang Liu, Bin Xia +4 authors
Magic Mirror framework uses Video Diffusion Transformers to generate identity-preserved videos with natural motion and high quality, overcoming limitations of existing methods.
Nandan Kumar Jha, Brandon Reagen
An information-theoretic framework is introduced to optimize transformer architectures for private inference by addressing the impact of nonlinearities on entropy and attention mechanisms.
Yekun Ke, Xiaoyu Li, Yingyu Liang +3 authors
Analysis of Visual Autoregressive (VAR) models identifies conditions for sub-quadratic time complexity and presents efficient constructions using low-rank approximations, advancing scalable image generation.
Hieu Man, Nghia Trung Ngo, Viet Dac Lai +3 authors
LUSIFER, a zero-shot approach using a multilingual encoder and LLM-based embedding model, enhances multilingual performance in text embedding tasks without multilingual supervision.
Mingyang Song, Zhaochen Su, Xiaoye Qu +2 authors
PRMBench is a process-level benchmark designed to evaluate the fine-grained error detection capabilities of Process-Level Reward Models (PRMs) across multiple dimensions.
Chengxing Xie, Bowen Li, Chang Gao +4 authors
SWE-Fixer is an open-source Large Language Model designed to resolve GitHub issues using two modules: a code file retrieval module that uses BM25 and a lightweight LLM, and a code editing module that generates patches for identified files, achieving state-of-the-art performance on the SWE-Bench Lite and Verified benchmarks.
Junjie Ye, Zhengyin Du, Xuesong Yao +11 authors
ToolHop, a new dataset for evaluating multi-hop tool use in large language models, highlights significant challenges and variations in tool-use strategies across different model families.
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