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。
44 篇论文 · 按点赞排序
Daniele Reda, Jungdam Won, Yuting Ye +2 authors
Real-time motion retargeting from sparse human sensor data to diverse avatar morphologies is achieved using reinforcement learning in physics simulators.
Jakob Drachmann Havtorn, Amelie Royer, Tijmen Blankevoort +1 authors
A dynamic mixed-scale tokenization scheme for Vision Transformers adjusts token scale based on image content, improving accuracy and computational efficiency.
Meysam Alizadeh, Maël Kubli, Zeynab Samei +4 authors
Open-source LLMs demonstrate competitive performance in text annotation tasks against proprietary models and human-based services.
Yihong Chen, Kelly Marchisio, Roberta Raileanu +4 authors
Active forgetting mechanism during pretraining improves pretrained language models' adaptation to new, especially distant languages with limited data.
Aaron Mueller, Kanika Narang, Lambert Mathias +2 authors
Meta-training with demonstration retrieval enhances few-shot NLP performance by leveraging dense passage retrievers and external knowledge sources.
Vivek Myers, Andre He, Kuan Fang +7 authors
A method is introduced to enable robots to follow natural language instructions with minimal labeled data by aligning language to the change between images rather than the goal image, enhancing policy training and generalization.
Feifan Song, Bowen Yu, Minghao Li +4 authors
A new method, Preference Ranking Optimization (PRO), is introduced to align LLMs with human preferences by extending the Bradley-Terry model to longer preference rankings, outperforming existing reinforcement learning techniques.
Liangzhe Yuan, Nitesh Bharadwaj Gundavarapu, Long Zhao +14 authors
VideoGLUE score measures foundation models' adaptation to various video tasks, revealing the superiority of video-native models over task-specialized and image-native models under different adaptions.
Yueh-Hua Wu, Xiaolong Wang, Masashi Hamaya
Elastic Decision Transformer improves upon Decision Transformer by facilitating trajectory stitching through dynamic history adjustment, achieving performance on par with Q Learning in multi-task scenarios.
Chenfei Kang, Peiling Lu, Botao Yu +4 authors
EmoGen generates emotional music using a two-stage process that translates emotion labels to general music attributes and then generates music via self-supervised learning, overcoming subjective bias.
Zhenyu Wu, Ziwei Wang, Xiuwei Xu +2 authors
TaPA integrates LLMs with visual perception to generate feasible and effective plans for embodied agents in complex environments, outperforming other models like LLaVA and GPT-3.5.
Zhongjin Luo, Dong Du, Heming Zhu +3 authors
A sketch-based system called SketchMetaFace enables amateur users to model high-fidelity 3D faces efficiently by using curvature-aware strokes and a novel method called Implicit and Depth Guided Mesh Modeling (IDGMM).
Lorenzo Noci, Chuning Li, Mufan Bill Li +4 authors
A modified Softmax-based attention model with skip connections exhibits well-behaved covariance structure and stability in the proportional limit of infinite-depth-and-width, preventing rank degeneracy issues.
Shuyang Sun, Weijun Wang, Qihang Yu +3 authors
ReMaX introduces relaxation techniques to improve mask transformer training for panoptic segmentation without extra inference costs, achieving state-of-the-art results with efficient backbones on benchmarks like COCO, ADE20K, and Cityscapes.
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