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
Renjie Li, Ruijie Ye, Mingyang Wu +4 authors
A large-scale benchmark, MMHU, is proposed for human behavior analysis in autonomous driving, featuring rich annotations and diverse data sources, and benchmarking various tasks including motion prediction and behavior question answering.
Zhuoshi Pan, Qizhi Pei, Yu Li +5 authors
REST, a stress-testing framework, evaluates large reasoning models under simultaneous multi-problem conditions, revealing performance differences and insights into model behavior under real-world reasoning demands.
Pengfei Zhou, Jie Xia, Xiaopeng Peng +14 authors
LoongX, a hands-free image editing system, uses multimodal neurophysiological signals and diffusion models to achieve performance comparable to text-driven methods and outperforms them when combined with speech.
Shaocong Dong, Lihe Ding, Xiao Chen +10 authors
CoPart, a part-aware diffusion framework, enhances 3D generation by decomposing objects into contextual part latents, improving detail, part relationships, and controllability.
Qiang Wang, Mengchao Wang, Fan Jiang +3 authors
FantasyPortrait, a diffusion transformer framework, generates high-fidelity and emotion-rich facial animations for single and multi-character scenarios using implicit representations and a masked cross-attention mechanism.
Shuyang Xu, Zhiyang Dou, Mingyi Shi +8 authors
A diffusion-based generative framework, MOSPA, is introduced to model human motion in response to spatial audio, achieving state-of-the-art performance using the newly created SAM dataset.
Chenguo Lin, Yuchen Lin, Panwang Pan +4 authors
MoVieS synthesizes 4D dynamic novel views from monocular videos using pixel-aligned Gaussian primitives, enabling unified appearance, geometry, and motion modeling within a single framework.
Jingze Zhu, Yongliang Wu, Wenbo Zhu +7 authors
A token-aware, layer-localized contrastive decoding method improves factual generation in large language models by selectively suppressing attention to specific token types at their respective depths.
Yilun Zhao, Weiyuan Chen, Zhijian Xu +5 authors
AbGen evaluates LLMs in designing ablation studies for scientific research, revealing performance gaps compared to human experts and highlighting the unreliability of current automated evaluation methods.
Chien Van Nguyen, Ruiyi Zhang, Hanieh Deilamsalehy +9 authors
Lizard is a linearization framework that transforms Transformer-based LLMs into subquadratic architectures for efficient infinite-context generation, using a hybrid attention mechanism and hardware-aware training.
Wasi Uddin Ahmad, Somshubra Majumdar, Aleksander Ficek +6 authors
OpenCodeReasoning-II, a large dataset for code reasoning, enhances code generation and critique through a two-stage fine-tuning strategy, improving competitive coding performance and extending LiveCodeBench for C++.
Seokhee Hong, Sunkyoung Kim, Guijin Son +3 authors
Two Korean expert-level benchmarks, KMMLU-Redux and KMMLU-Pro, are introduced to evaluate Large Language Models in industrial and professional contexts.
Taolin Zhang, Maosong Cao, Alexander Lam +2 authors
CompassJudger-2, a generalist judge model, improves cross-domain evaluation accuracy and robustness through task-driven data curation and a refined learning objective.
Yuxi Xiao, Jianyuan Wang, Nan Xue +7 authors
SpatialTrackerV2 is a feed-forward 3D point tracking method for monocular videos that integrates scene geometry, camera ego-motion, and object motion into a unified, differentiable architecture, achieving high accuracy and speed.
Corrado Rainone, Tim Bakker, Roland Memisevic
A new approach formats tokens as a multi-turn interaction trace with a stateful tool for training Large Language Models, enabling faster sampling and denser reward signals for tasks like repairing Python code.
Arushi Goel, Sreyan Ghosh, Jaehyeon Kim +8 authors
Audio Flamingo 3, a state-of-the-art audio-language model, advances reasoning and understanding across speech, sound, and music through a unified audio encoder and novel training strategies.
Florian Grötschla, Luis Müller, Jan Tönshoff +2 authors
AgentsNet is a new benchmark for evaluating multi-agent systems' ability to self-organize, communicate, and solve problems collaboratively across varying network sizes.
Seungwoo Kim, Khai Loong Aw, Klemen Kotar +8 authors
Counterfactual prompting of generative video models, using KL-tracing, extracts optical flow without fine-tuning, outperforming state-of-the-art methods on real and synthetic datasets.
Ziye Li, Hao Luo, Xincheng Shuai +1 authors
AnyI2V is a training-free framework that animates conditional images with user-defined motion trajectories, supporting various data types and enabling flexible video generation.
Yilun Zhao, Chengye Wang, Chuhan Li +1 authors
A benchmark evaluates multimodal models' ability to interpret schematic diagrams in scientific literature, revealing performance gaps compared to human experts.
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