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
Mathieu Andreux, Märt Bakler, Yanael Barbier +50 authors
Surfer 2, a unified visual-based architecture, achieves state-of-the-art performance across web, desktop, and mobile environments without task-specific fine-tuning, demonstrating the potential of systematic orchestration for general-purpose computer control.
Minho Park, Kinam Kim, Junha Hyung +5 authors
Action Coherence Guidance (ACG) improves action coherence in Vision-Language-Action (VLA) models during test time, enhancing performance in diverse manipulation tasks.
Shengnan An, Xunliang Cai, Xuezhi Cao +8 authors
AMO-Bench is a benchmark for evaluating advanced mathematical reasoning in large language models with problems at the International Mathematical Olympiad level or higher, revealing significant room for improvement in current models.
Dan Hendrycks, Dawn Song, Christian Szegedy +30 authors
A quantifiable framework based on Cattell-Horn-Carroll theory evaluates AI systems across ten cognitive domains, revealing significant gaps in foundational cognitive abilities like long-term memory.
Ziyu Guo, Xinyan Chen, Renrui Zhang +7 authors
Video models like Veo-3 show promise in short-term visual reasoning but struggle with long-term causal reasoning and abstract logic, indicating they are not yet reliable standalone zero-shot reasoners.
Baolu Li, Yiming Zhang, Qinghe Wang +8 authors
VFXMaster is a unified framework for generating VFX effects using in-context learning, enabling generalization to unseen effects and efficient adaptation from a single video.
Yueqi Song, Ketan Ramaneti, Zaid Sheikh +18 authors
The agent data protocol (ADP) standardizes diverse agent training datasets, enabling improved performance across various tasks without domain-specific tuning.
Qi Liu, Yanzhao Zhang, Mingxin Li +3 authors
A unified framework extends a single text embedding model to perform both retrieval and listwise reranking, achieving state-of-the-art results with low latency.
Zhanchao Zhou, Xiaodong Chen, Haoxing Chen +2 authors
Knocking-heads attention (KHA) enhances multi-head attention by enabling cross-head interactions, improving training dynamics and performance in large language models.
Yatai Ji, Teng Wang, Yuying Ge +4 authors
ReDiff, a refining-enhanced diffusion framework, addresses train-inference discrepancies in discrete diffusion models by enabling the model to identify and correct its own errors, improving coherence and factual accuracy in generated content.
Zewen Chi, Li Dong, Qingxiu Dong +4 authors
AsyncThink, a new reasoning paradigm for large language models, improves inference efficiency and accuracy through concurrent and optimized sub-query processing.
Pengtao Chen, Xianfang Zeng, Maosen Zhao +7 authors
RegionE is a region-aware framework that accelerates instruction-based image editing by differentiating between edited and unedited regions, optimizing the denoising process for each.
Zhuoran Jin, Hongbang Yuan, Kejian Zhu +5 authors
Omni-Reward addresses modality imbalance and preference rigidity in reward models by introducing a benchmark, dataset, and model that support multiple modalities and free-form preferences.
Zhengshen Zhang, Hao Li, Yalun Dai +10 authors
FALCON enhances vision-language-action models by integrating rich 3D spatial tokens into the action head, improving spatial reasoning and modality transferability.
Jing Lin, Ruisi Wang, Junzhe Lu +10 authors
A framework transfers knowledge from video generation to 3D human motion generation using a large dataset, a flow-matching-based diffusion transformer, and a hierarchical benchmark, improving generalization and performance.
Xuanpu Zhang, Xuesong Niu, Ruidong Chen +6 authors
Group Relative Attention Guidance enhances image editing quality by modulating token deltas in Diffusion-in-Transformer models, providing fine-grained control over editing intensity.
Yuqian Yuan, Wenqiao Zhang, Xin Li +6 authors
PixelRefer is a unified region-level multimodal large language model framework that enhances fine-grained object-centric understanding and efficiency through a Scale-Adaptive Object Tokenizer and Object-Centric Infusion module.
Yida Zhao, Kuan Li, Xixi Wu +11 authors
Entity-aware Group Relative Policy Optimization (E-GRPO) enhances search agents by incorporating entity information into the reward function, improving accuracy and efficiency in complex, knowledge-intensive tasks.
Ranran Haoran Zhang, Soumik Dey, Ashirbad Mishra +3 authors
Batch speculative decoding improves LLM inference throughput by managing ragged tensors to maintain output equivalence and reduce realignment overhead.
Guoxin Chen, Jing Wu, Xinjie Chen +6 authors
ReForm, a reflective autoformalization method, improves the semantic accuracy of formal statements generated from natural language mathematics through iterative refinement and semantic consistency evaluation.
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