图像解码器；多头混合专家网络；视觉模型美学对齐；医学视觉任务适应基准
编者按：欢迎阅读“科研上新”栏目！“科研上新”汇聚了微软亚洲研究院最新的创新成果与科研动态。在这里，你可以快速浏览研究院的亮点资讯，保持对前沿领域的敏锐嗅觉，同时也能找到先进实用的开源工具。 论文链接：https://arxiv.org/abs/2404.12876 (opens in new tab) 近年来，深度学习的显著进步极大地推动了计算机视觉领域的发展，尤其是视觉 Transformer...
In the news | Microsoft on the Issues
Protecting the public from abusive AI-generated content
AI-generated deepfakes are realistic, easy for nearly anyone to make, and increasingly being used for fraud, abuse, and manipulation – especially to target kids and seniors. While the tech sector and non-profit groups have taken recent steps to address this…
Abstracts: July 29, 2024
| Gretchen Huizinga and Li Lyna Zhang
A lack of appropriate data, decreased model performance, and other obstacles have made it difficult to expand the input language models can receive. Li Lyna Zhang introduces LongRoPE, a method capable of extending content windows to more than 2 million…
Awards | ACM SIGIR
Bhaskar Mitra receives two ACM SIGIR Early Career Researcher Awards
Bhaskar Mitra recently received the ACM SIGIR Early Career Researcher Awards for the following two categories: Excellence in Research: "For high-impact work, including research in neural IR and the establishing the MS-MARCO ranking benchmark" and Excellence in Community Engagement: "For…
In the news | The Stack
‘Enormous business potential’: Microsoft on why GraphRAG outperforms naive RAG
Redmond opens up to discuss its new tool, which can extract data from unstructured text using large language models. Microsoft's GraphRAG is a new approach to Retrieval-Augmented Generation (RAG) that Redmond has described a "significant advance in enhancing the capability…
大语言模型应用如何实现端到端优化？
作者：系统组（上海） 编者按：基于大语言模型（LLMs）开发的应用目前主要使用公共 LLMs 服务提供的 API 进行，但是这些 LLMs 服务的 API 设计以请求为中心，缺乏应用级信息，难以有效优化整个应用流程，影响任务的端到端性能。为此，微软亚洲研究院的研究员们开发了一个专注于 LLMs 应用端到端体验的服务系统 Parrot，它具有减少网络延迟、提高吞吐量、减少冗余计算等优势。Parrot...
Tracing the path to self-adapting AI agents
| Ching-An Cheng, Adith Swaminathan, and Allen Nie
Introducing Trace, Microsoft and Stanford University's novel AI optimization framework, now available as a Python library. Trace adapts dynamically and optimizes a wide range of applications from language models to robot control.
Awards | Stealing Part of a Production Language Model
ICML 2024 Best Paper Award
"Stealing Part of a Production Language Model" written by A. Feder Cooper et al. receives ICML 2024 Best Paper Award.
LSR-MP：突破分子模拟瓶颈，探索知识引导的AI范式
作者：科学智能中心 编者按：分子动力学模拟在新药开发、材料设计等领域发挥着重要作用。近年来机器学习技术的不断发展，使得其对分子间相互作用的刻画也更加精确，但却面临着随分子体系扩大，计算效率降低和长程信息丢失的难题。在此背景下，微软研究院科学智能中心同耶鲁大学、西安交通大学提出了一种名为 LSR-MP 的新型分子动力学机器学习框架。该框架结合了物理洞见和几何深度学习，通过在原子/分子片段上分别建模短...