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
Matthew Chang, Theophile Gervet, Mukul Khanna +10 authors
GOAT is a universal navigation system for mobile robots that uses multimodal input, learns from its experience, and is adaptable across different robotic platforms, achieving high success rates in navigation and other tasks.
Ziyi Lin, Chris Liu, Renrui Zhang +13 authors
SPHINX, a multi-modal large language model, enhances vision-language alignment and multi-purpose capabilities through joint weight and task mixing, resulting in superior performance across various applications and image parsing tasks.
An Yan, Zhengyuan Yang, Wanrong Zhu +9 authors
A GPT-4V-based agent demonstrates high accuracy in zero-shot smartphone GUI navigation through advanced screen interpretation, action reasoning, and precise localization.
Daniel Y. Fu, Hermann Kumbong, Eric Nguyen +1 authors
FlashFFTConv optimizes FFT convolutions by leveraging matrix multiply units and sparse techniques, improving speed and model performance across various tasks including long-sequence DNA processing and high-resolution vision.
Keming Lu, Hongyi Yuan, Runji Lin +4 authors
Zooter, a reward-guided routing method for Large Language Models, improves efficiency and performance by precisely directing queries to the most suitable model, outperforming existing ensemble techniques.
Jorge Leandro, Sudha Rao, Michael Xu +4 authors
A graph-based interactive narrative visualization system called GRIM uses GPT-4 to generate branching narratives efficiently with designer constraints.
Sanchit Ahuja, Divyanshu Aggarwal, Varun Gumma +8 authors
A study expands the MEGA benchmarking suite to MEGAVERSE with 22 datasets covering 81 languages, evaluating various LLMs including GPT-3.5-Turbo, GPT4, PaLM2, Llama2, and LLaVa-v1.5, highlighting performance gaps between models on low-resource languages.
Microsoft Research AI4Science, Microsoft Azure Quantum
GPT-4 shows promising potential across various scientific applications, including drug discovery, biology, computational chemistry, materials design, and PDE, by demonstrating comprehension and problem-solving capabilities.
Wenting Zhao, Justin T Chiu, Jena D. Hwang +6 authors
Research focuses on uncommonsense abductive reasoning using a new corpus to improve model explanations of unexpected outcomes, showing that model-augmented human explanations and online imitation learning enhance performance.
Kumar Shridhar, Koustuv Sinha, Andrew Cohen +6 authors
The ART method improves the refinement of Large Language Model (LLM) outputs on multistep reasoning tasks by using smaller models to decide when and how to refine the initial predictions.
Vasilisa Bashlovkina, Zhaobin Kuang, Riley Matthews +4 authors
A new evaluation metric called trusted source alignment (TSA) is introduced and measured for large language models to assess their preference for content from trusted publishers over contradicting information.
Jiawei Lin, Jiaqi Guo, Shizhao Sun +3 authors
LayoutPrompter uses large language models and in-context learning to generate high-quality layouts with versatility and data efficiency.
Archiki Prasad, Alexander Koller, Mareike Hartmann +4 authors
ADaPT, a new approach for complex tasks, uses as-needed decomposition and planning, outperforming baselines in various interactive decision-making tasks.
Yuliang Liu, Xiangru Tang, Zefan Cai +23 authors
ML-Bench and ML-Agent evaluate and enhance the ability of large language models to use existing open-source libraries for machine learning tasks.
AJ Piergiovanni, Isaac Nobel, Dahun Kim +3 authors
Mirasol3B, a multimodal model with separate autoregressive components for time-synchronized and context modalities, achieves state-of-the-art performance by learning compact representations and managing high computational demands.
Yun Zhu, Nevan Wichers, Chu-Cheng Lin +8 authors
SiRA, a sparse mixture of low-rank adaptation with expert dropout and capacity limits, improves performance over LoRA and other mixture of expert methods in various tasks.
Yunfei Chu, Jin Xu, Xiaohuan Zhou +5 authors
Qwen-Audio addresses the limitation of diverse audio types and tasks in instruction-following models through a multi-task training framework with hierarchical tagging, achieving superior performance across benchmarks and enabling Qwen-Audio-Chat for versatile dialogue scenarios.
Suyu Ge, Chunting Zhou, Rui Hou +5 authors
A multi-round automatic red-teaming method uses iterative adversarial prompts and safety fine-tuning to enhance the safety of large language models with minimal impact on their helpfulness.
Sheshera Mysore, Zhuoran Lu, Mengting Wan +6 authors
PEARL, a retrieval-augmented writing assistant, personalizes large language model outputs for users by selecting relevant documents, improving the quality of workplace social media posts and Reddit comments.
Mahmoud G. Salem, Jiayu Ye, Chu-Cheng Lin +1 authors
Unsupervised pretraining of non-autoregressive T5 models using unrolled denoising achieves state-of-the-art results in downstream generation tasks.
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