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
Junda Wu, Zachary Novack, Amit Namburi +5 authors
FUTGA enhances music captioning by generating detailed, temporally-aware descriptions using generative augmentation with temporal compositions and large language models.
Nitzan Bitton-Guetta, Aviv Slobodkin, Aviya Maimon +6 authors
Visual Riddles is a benchmark that tests vision and language models on visual riddles requiring commonsense and world knowledge, highlighting the gap between human and model performance.
Xi Victoria Lin, Akshat Shrivastava, Liang Luo +5 authors
MoMa, an architecture for pre-training mixed-modal language models, achieves substantial efficiency savings by using modality-specific experts and mixture-of-depths, though combining these techniques affects causal inference performance.
Francis Kulumba, Wissam Antoun, Guillaume Vimont +1 authors
HALvest bridges citation networks with HAL documents, creating a multilingual dataset for training language models, authorship attribution, and graph representation learning.
Benjamin Clavié
Improvements to the training and inference methods of multi-vector models like JaColBERT enhance performance in Japanese information retrieval, achieving state-of-the-art results.
Penglei Sun, Yaoxian Song, Xiang Liu +5 authors
The research introduces a 3D multimodal question answering dataset and method for city-level scene understanding, incorporating spatial semantics and human-environment interactions, achieving state-of-the-art performance.
Tianhao Wu, Weizhe Yuan, Olga Golovneva +5 authors
A novel Meta-Rewarding step in self-improvement enhances LLM judgment capabilities and instruction-following ability without human supervision.
Mohammed Khalil, Mohammed Sabry
The creation and evaluation of the ATHAR dataset, containing 66,000 high-quality Classical Arabic to English translation samples, demonstrates the need for extensive datasets to improve current large language models.
Jun Wang, Ying Yuan, Haichuan Che +4 authors
A reinforcement learning-based approach trains a sensorimotor policy for in-hand manipulation of pen-like objects through simulation and real-world fine-tuning with limited trajectories.
Delyan Boychev, Radostin Cholakov
ImagiNet, a high-resolution dataset, aids in synthetic image detection using ResNet-50 trained with SelfCon, achieving high accuracy under various conditions.
Yunze Man, Yichen Sheng, Jianming Zhang +2 authors
A novel method ORG reconstructs 3D object geometry and ground surface to improve shadow rendering and pose manipulation in 3D-aware image editing tasks.
Yuheng Shi, Minjing Dong, Mingjia Li +1 authors
The Visual State Space Duality (VSSD) model enhances the performance and efficiency of State Space Models in vision tasks by introducing a non-causal format.
Shanbo Cheng, Zhichao Huang, Tom Ko +4 authors
CLASI, a high-quality Simultaneous Speech Translation system, utilizes a data-driven strategy and multi-modal retrieving module to achieve superior translation quality compared to existing systems in various challenging scenarios.
Giulio Corallo, Paolo Papotti
Finch enables large input processing in models by compressing context via pre-trained self-attention weights, maintaining semantic integrity without fine-tuning.
Gilad Deutch, Rinon Gal, Daniel Garibi +2 authors
Methods are proposed to enhance fast sampling in text-based diffusion models for image editing, addressing visual artifacts and insufficient editing strength.
Wenjun Zeng, Yuchi Liu, Ryan Mullins +9 authors
ShieldGemma, a suite of LLM-based safety content moderation models, outperforms existing solutions across various harm types and demonstrates strong generalization even when trained on synthetic data.
Manuel Kansy, Jacek Naruniec, Christopher Schroers +2 authors
Motion-textual inversion uses a pre-trained image-to-video model and text/image embeddings to transfer motion from a reference video to different images or objects, achieving high-quality and semantically similar motion transfers.
Weihao Yu, Zhengyuan Yang, Linfeng Ren +7 authors
MM-Vet v2 introduces a new benchmark for evaluating multimodal models by including image-text sequence understanding, expanding the evaluation set size, and identifying Claude 3.5 Sonnet as the top-performing model.
Brigita Vileikytė, Mantas Lukoševičius, Lukas Stankevičius
Transformer models, particularly BERT and T5, perform well in sentiment analysis of Lithuanian online reviews, surpassing commercial state-of-the-art models.
ShahRukh Athar, Shunsuke Saito, Zhengyu Yang +2 authors
A method is proposed to generate studio-quality facial texture maps from monocular phone captures using StyleGAN2 and diffusion models, enabling photorealistic avatar creation.
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