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
Kling Team, Jialu Chen, Yikang Ding +21 authors
Kling-MotionControl is a DiT-based framework for character animation that combines heterogeneous motion representations, adaptive identity-agnostic learning, and advanced acceleration techniques to achieve high-fidelity, expressive, and controllable video generation.
Marco Garosi, Matteo Farina, Alessandro Conti +2 authors
Large Multimodal Models demonstrate superior performance in closed-world classification with in-context learning and excel in open-world scenarios when equipped with context refinement techniques.
Lulu Hu, Wenhu Xiao, Xin Chen +4 authors
Post-training quantization for multimodal large language models addresses smoothing misalignment and cross-modal computational invariance through modality-aware smoothing and cross-modal compensation techniques.
Ziwen Xu, Kewei Xu, Haoming Xu +8 authors
SteerEval is a hierarchical benchmark for evaluating large language model controllability across language features, sentiment, and personality domains with three specification levels.
Zihao Huang, Tianqi Liu, Zhaoxi Chen +7 authors
ArtHOI synthesizes articulated human-object interactions by formulating 4D reconstruction from monocular video priors, using optical flow for part segmentation and a decoupled reconstruction pipeline to achieve physically plausible and geometrically consistent interactions.
Yiqi Lin, Guoqiang Liang, Ziyun Zeng +3 authors
A scalable data generation pipeline creates high-fidelity video editing training data, and a unified architecture enables improved instruction-following and reference fidelity in controllable video editing.
Jinpeng Chen, Cheng Gong, Hanbo Li +9 authors
CoVe is a post-training data synthesis framework that generates high-quality training trajectories for interactive tool-use agents by incorporating task constraints as verification mechanisms, achieving superior performance compared to baseline models.
Zebin You, Xiaolu Zhang, Jun Zhou +2 authors
LLaDA-o is an omni diffusion model that uses a Mixture of Diffusion framework to jointly handle text understanding and visual generation through a shared attention backbone, achieving state-of-the-art performance in multimodal tasks.
Yong Liu, Xingjian Su, Shiyu Wang +7 authors
Timer-S1 is a scalable Mixture-of-Experts time series model with 8.3B parameters that uses serial scaling and novel TimeMoE blocks to improve long-term forecasting accuracy.
Jyoti Aneja, Michael Harrison, Neel Joshi +3 authors
A compact open-weight multimodal reasoning model is presented that achieves competitive performance through careful architecture design, high-quality data curation, and a hybrid approach combining direct answering with chain-of-thought reasoning.
George Bredis, Nikita Balagansky, Daniil Gavrilov +1 authors
NE-Dreamer uses a temporal transformer to predict next-step encoder embeddings for model-based reinforcement learning without requiring decoders or auxiliary supervision.
Rituraj Sharma, Weiyuan Chen, Noah Provenzano +1 authors
PRISM is a Process Reward Model-guided inference algorithm that enhances DEEPTHINK systems by using step-level verification to improve population refinement and solution aggregation, achieving strong performance on mathematical and scientific benchmarks.
Baoqing Yue, Zihan Zhu, Yifan Zhang +3 authors
Interactive benchmarks offer a unified framework for evaluating model intelligence through active information acquisition under constraint conditions, demonstrating superior assessment of reasoning capabilities compared to traditional benchmarks.
Zhenting Wang, Huancheng Chen, Jiayun Wang +1 authors
A memory mechanism called Memex enables large language model agents to handle long-horizon tasks more effectively by maintaining compact context through structured summaries while storing full interaction details in an external database, allowing selective retrieval based on learned criteria.
Jintao Zhang, Marco Chen, Haoxu Wang +5 authors
Research investigates why low-bit attention methods like SageBwd exhibit performance gaps during pre-training and identifies key factors for stable training with reduced precision.
Dadi Guo, Yuejin Xie, Qingyu Liu +7 authors
Code agents can autonomously generate more complex mathematical problems by evolving existing ones, providing a scalable solution for creating high-difficulty reasoning problems.
Alexander Samarin, Sergei Krutikov, Anton Shevtsov +3 authors
Speculative decoding speedup is improved by LK losses that directly optimize acceptance rate instead of KL divergence, showing consistent performance gains across different model sizes and domains.
Zhengqing Yuan, Kaiwen Shi, Zheyuan Zhang +3 authors
A comprehensive benchmark and detection framework for identifying fabricated citations in scientific writing using a multi-agent verification pipeline that assesses citation faithfulness and evidence alignment.
Francesco Laiti, Davide Talon, Jacopo Staiano +1 authors
Memorability Feedback enables users to improve photo memorability through natural language guidance by leveraging multimodal large language models and a teacher-student steering strategy.
Yisu Zhang, Chenjie Cao, Tengfei Wang +4 authors
WorldStereo enables camera-controlled video generation and 3D reconstruction through geometric memory modules that provide structural priors and 3D correspondence constraints.
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