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
Imanol Schlag, Sainbayar Sukhbaatar, Asli Celikyilmaz +4 authors
Embedding large pre-trained language models within algorithms improves question-answering performance by 6.4% compared to chain-of-thought methods without fine-tuning.
Anastasia Razdaibiedina, Yuning Mao, Rui Hou +4 authors
Residual Prompt Tuning enhances prompt tuning for parameter-efficient pre-trained language models, improving performance and stability while reducing prompt length.
Feilong Chen, Minglun Han, Haozhi Zhao +4 authors
X-LLM integrates multimodal information into language models using X2L interfaces, demonstrating multimodal chat abilities competitive with GPT-4.
Xin-Yang Zheng, Hao Pan, Peng-Shuai Wang +3 authors
A diffusion-based framework offering improved local controllability and generalizability for 3D shape generation from 2D sketches.
Zhengyi Luo, Jinkun Cao, Alexander Winkler +2 authors
A physics-based humanoid controller imitates motion and recovers from faults using a progressive multiplicative control policy, demonstrating high fidelity and scalability in noisy and dynamic environments.
Suzanne Petryk, Spencer Whitehead, Joseph E. Gonzalez +3 authors
Token-Level Confidence (TLC) improves caption correctness by assessing fine-grained details, reducing errors like object hallucination and enhancing compositional reasoning.
Haoyang Huang, Tianyi Tang, Dongdong Zhang +4 authors
Cross-lingual-thought prompting improves multilingual performance and reduces performance gaps across different tasks and languages.
Yi Dai, Hao Lang, Yinhe Zheng +3 authors
Diana is a dynamic architecture-based lifelong learning model using prompt-enhanced language models to handle a sequence of tasks, including unseen tasks, with improved generalization and performance.
Wang-Cheng Kang, Jianmo Ni, Nikhil Mehta +4 authors
LLMs show data-efficient potential in user rating prediction through fine-tuning, though they lag behind CF methods in zero-shot scenarios.
Wen-Yu Hua, Brian Williams, Davood Shamsi
LACoS-BLOOM, an 8-bit Low-rank Adaptation with Contrastive objective of Siamese-BLOOM, improves text embeddings quality and efficiency in NLP tasks.
Kun Su, Judith Yue Li, Qingqing Huang +8 authors
A novel multi-stage autoregressive model, V2Meow, generates high-quality music audio aligned with various video inputs, outperforming existing methods in visual-audio correspondence and audio quality.
Mustafa Işık, Martin Rünz, Markos Georgopoulos +4 authors
HumanRF is a 4D neural scene representation that captures high-fidelity, full-body motion for novel view synthesis, utilizing a temporal matrix-vector decomposition for high-resolution data.
Gengshan Yang, Chaoyang Wang, N Dinesh Reddy +1 authors
RAC learns animatable 3D models from monocular videos by optimizing instance-specific skeletons, regularizing latent spaces, and disentangling objects from backgrounds, enabling category-wide models.
Chen Bao, Helin Xu, Yuzhe Qin +1 authors
A new benchmark, DexArt, is proposed to evaluate the generalizability of reinforcement learning policies in dexterous manipulation of articulated objects using a multi-finger robot hand in a physical simulator.
Hassan Akbari, Dan Kondratyuk, Yin Cui +3 authors
Integrated Multimodal Perception (IMP) combines Alternating Gradient Descent and Mixture-of-Experts in a single Transformer encoder to improve multimodal understanding and achieve state-of-the-art zero-shot video classification.
Naveen Ram, Dima Kuzmin, Ellie Ka In Chio +4 authors
A unified transformer model, fine-tuned on conversational recommendations with additional multitask learning, achieves competitive performance in both item recommendation and dialogue generation.
Aldo Gael Carranza, Rezsa Farahani, Natalia Ponomareva +3 authors
A method using differentially private large language models for generating private synthetic queries enhances the security and effectiveness of deep retrieval models in recommender systems.
Andrea Burns, Krishna Srinivasan, Joshua Ainslie +5 authors
WikiWeb2M is introduced to provide a comprehensive dataset containing images, text, and structure data from Wikipedia for multimodal webpage understanding tasks.
Zerong Zheng, Xiaochen Zhao, Hongwen Zhang +2 authors
A new method called AvatarReX enables real-time rendering of expressive full-body avatars from video data using a compositional avatar representation and a deferred rendering pipeline.
Ze-Xin Yin, Jiaxiong Qiu, Ming-Ming Cheng +1 authors
A multi-space neural radiance field (MS-NeRF) enhances rendering quality for scenes with reflective and refractive objects by using multiple feature fields in parallel sub-spaces.
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