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
Haonan Qiu, Menghan Xia, Yong Zhang +4 authors
A method called FreeNoise is introduced to enhance video diffusion models for generating high-fidelity, long videos conditioned on multiple text prompts with minimal time cost.
Qianchu Liu, Stephanie Hyland, Shruthi Bannur +16 authors
GPT-4 performs effectively on various radiology tasks, matching or outperforming state-of-the-art radiology-specific models with prompting strategies.
Haoran Li, Yiran Liu, Xingxing Zhang +2 authors
A novel finetuning method using probabilistic and contextual ranking improves instruction-tuned LLM performance across multiple benchmarks by leveraging stronger teacher models.
Mengdi Xu, Peide Huang, Wenhao Yu +7 authors
RoboTool, a system utilizing Large Language Models, enables robots to creatively use tools by interpreting natural language and generating executable code for complex tasks involving implicit physical constraints.
Xavier Puig, Eric Undersander, Andrew Szot +20 authors
Habitat 3.0 provides a simulation platform for studying collaborative human-robot tasks with accurate humanoid simulation, human-in-the-loop infrastructure, and support for social navigation and rearrangement tasks.
Muhammad Ferjad Naeem, Yongqin Xian, Xiaohua Zhai +3 authors
Adding local-to-global correspondence learning via self-distillation to contrastive pre-training improves model performance across various computer vision tasks, particularly segmentation, and sets new benchmarks in zero-shot classification, few-shot classification, and retrieval.
Arno Candel, Jon McKinney, Philipp Singer +4 authors
An open-source ecosystem is introduced for developing and testing Large Language Models (LLMs), including a family of fine-tuned models and a GUI for efficient use of state-of-the-art techniques.
Nicklas Hansen, Hao Su, Xiaolong Wang
TD-MPC2, an improved model-based reinforcement learning algorithm, achieves strong performance across various RL tasks with consistent results using a single set of hyperparameters and benefits from larger models and datasets.
Swarnadeep Saha, Omer Levy, Asli Celikyilmaz +3 authors
The Branch-Solve-Merge (BSM) method enhances Large Language Models for complex tasks by decomposing them into sub-tasks, solving each independently, and then merging the results, improving evaluation and constrained text generation.
Tianshi Cao, Karsten Kreis, Sanja Fidler +2 authors
TexFusion uses text-guided latent diffusion models to synthesize 3D textures from 2D renders, providing high-quality, globally coherent results without reliance on ground truth textures.
Samyadeep Basu, Nanxuan Zhao, Vlad Morariu +2 authors
Text-to-image diffusion models encode visual attributes across multiple components in the UNet and one causal state in the text-encoder, enabling efficient model editing with Diff-QuickFix.
Mrinank Sharma, Meg Tong, Tomasz Korbak +16 authors
RLHF-trained models frequently exhibit sycophantic behavior, which may be driven by human preferences favoring such responses.
Vikas Raunak, Hany Hassan Awadalla, Arul Menezes
Asymmetric perturbation of high-quality demonstrations in few-shot Machine Translation with LLMs shows that target-side perturbation significantly impacts translation quality, leading to the introduction of Zero-Shot-Context for enhanced zero-shot performance.
Xinyu Huang, Yi-Jie Huang, Youcai Zhang +6 authors
RAM++, a unified image-text interaction framework, enhances open-set recognition and predefined category identification by integrating large language models and semantic concepts into image tagging.
Aarush Gupta, Junli Cao, Chaoyang Wang +4 authors
Neural light field methods using light slab representation offer superior rendering quality and speed on mobile devices compared to previous approaches.
Marah I Abdin, Suriya Gunasekar, Varun Chandrasekaran +5 authors
KITAB, a new dataset, evaluates constraint satisfaction in large language models, highlighting their limitations in handling irrelevant information, factual errors, and incomplete responses, especially without context.
Young-Suk Lee, Md Arafat Sultan, Yousef El-Kurdi +4 authors
A new in-context learning method is proposed to generate high-quality instruction tuning data for smaller, open-access language models, improving their performance.
Zhongzhan Huang, Pan Zhou, Shuicheng Yan +1 authors
The paper analytically examines the instability of UNet in diffusion models due to long skip connections and proposes a coefficient scaling framework to enhance training stability and speed.
Sandipan Kundu, Yuntao Bai, Saurav Kadavath +33 authors
Constitutional AI using general principles can mitigate harmful behaviors in conversational models, reducing the need for detailed lists of principles.
Haofei Yu, Cunxiang wang, Yue Zhang +1 authors
A new memory selection strategy in the Transformer architecture enhances long-range language modeling by selecting relevant tokens based on a simple metric, improving performance without additional training or parameters.
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