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
Manel Baradad, Yuanzhen Li, Forrester Cole +4 authors
A novel Background Prompting strategy enhances depth estimation accuracy by adapting object images with learned backgrounds, improving focus and sim2real generalization.
Nan Liu, Yilun Du, Shuang Li +2 authors
An unsupervised method discovers generative concepts from images, enabling the creation of new images and use as a representation for classification tasks.
Quynh Phung, Songwei Ge, Jia-Bin Huang
Two novel losses improve attention focus in diffusion models, enhancing text-to-image alignment on benchmarks.
Jing Xu, Da Ju, Joshua Lane +10 authors
BlenderBot 3x, an updated conversational model, is trained using organic user data to enhance its skills and safety, showing preference in conversations and better handling of challenging situations.
Ye Lin, Xiaohui Wang, Zhexi Zhang +3 authors
MobileNMT system translates with low memory and fast decoding on mobile devices through model compression and INT8-friendly engine.
Zixin Zhu, Xuelu Feng, Dongdong Chen +5 authors
An asymmetric VQGAN enhances StableDiffusion by reducing information loss and improving inpainting and editing performance without altering the original model architecture.
Kaijie Zhu, Jindong Wang, Jiaheng Zhou +8 authors
PromptBench evaluates large language models' robustness to adversarial prompts across various tasks and datasets, revealing vulnerabilities and providing recommendations for prompt composition.
Gabriel Poesia, Kanishk Gandhi, Eric Zelikman +1 authors
A guide tool called LogicGuide improves language model accuracy in logical reasoning tasks by constraining generation to valid statements, reducing content effects and enabling self-improvement.
Hantian Ding, Varun Kumar, Yuchen Tian +9 authors
A static evaluation framework using Abstract Syntax Trees quantifies static errors in Python code completions, offering a more efficient alternative to execution-based evaluation for large models.
Zhaoshuo Li, Thomas Müller, Alex Evans +4 authors
Neuralangelo combines multi-resolution 3D hash grids with neural surface rendering for detailed 3D surface reconstruction from multi-view images without auxiliary inputs.
Zechun Liu, Barlas Oguz, Aasish Pappu +2 authors
Ternary and binary transformer models achieve significant efficiency gains with minor accuracy drops compared to full-precision counterparts, particularly in summarization and machine translation tasks.
Saurabh Saxena, Charles Herrmann, Junhwa Hur +4 authors
Denoising diffusion probabilistic models achieve state-of-the-art performance in optical flow and monocular depth estimation through self-supervised pre-training, handling of noisy data, and Monte Carlo inference.
Zeqiu Wu, Yushi Hu, Weijia Shi +6 authors
Fine-Grained Reinforcement Learning from Human Feedback improves language model performance by using detailed human feedback to provide segment-specific rewards for diverse feedback types.
Alexey Kurakin, Natalia Ponomareva, Umar Syed +2 authors
Using a properly tuned generative language model to create differentially private synthetic datasets achieves high-quality data for downstream tasks, including model tuning.
Zhizheng Zhang, Xiaoyi Zhang, Wenxuan Xie +1 authors
Responsible Task Automation (ResponsibleTA) leverages LLMs to ensure responsible collaboration between coordinators and executors in task automation by assessing feasibility, completeness, and security.
Carlos Esteves, Jean-Jacques Slotine, Ameesh Makadia
Scaling spherical CNNs achieves state-of-the-art performance in molecular benchmarks and competitive performance in weather forecasting by leveraging hardware accelerators and novel model components.
Duy-Kien Nguyen, Vaibhav Aggarwal, Yanghao Li +4 authors
R-MAE, an extension of Masked Autoencoding tailored for region-based pre-training, improves unsupervised learning in computer vision tasks with minimal overhead.
Shreyank N Gowda, Anurag Arnab, Jonathan Huang
A training strategy that freezes the spatial transformer in ViViT models reduces training costs and memory consumption while maintaining or improving performance.
Sanjay Subramanian, Medhini Narasimhan, Kushal Khangaonkar +6 authors
A framework generates modular Python code that combines pre-trained visual and language models for visual question answering, improving accuracy on benchmark datasets.
Haiyang Xu, Qinghao Ye, Xuan Wu +13 authors
A new large-scale Chinese video-language dataset (Youku-mPLUG) is released, enabling deep pre-training and evaluation of multimodal models. Researchers demonstrated significant improvements in video category classification and captioning using models pre-trained on this dataset.
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