AI在线 AI在线

Microsoft AI Unveils Code Researcher: 58% Crash Resolution Rate Stuns the Industry!

Microsoft AI has unveiled a groundbreaking tool called Code Researcher, designed specifically for handling large system code and commit history.. This innovative tool aims to tackle the challenges of debugging and fixing crashes in complex system codes, such as the Linux kernel, marking another significant breakthrough for AI in software development. According to the latest public information obtained by AIbase, Code Researcher enhances the efficiency and accuracy of system-level software maintenance through multi-step reasoning and semantic analysis.The Core Capabilities of Code Researcher.

Microsoft AI has unveiled a groundbreaking tool called Code Researcher, designed specifically for handling large system code and commit history.

This innovative tool aims to tackle the challenges of debugging and fixing crashes in complex system codes, such as the Linux kernel, marking another significant breakthrough for AI in software development. According to the latest public information obtained by AIbase, Code Researcher enhances the efficiency and accuracy of system-level software maintenance through multi-step reasoning and semantic analysis.

image.png

The Core Capabilities of Code Researcher

Code Researcher is an autonomous agent based on large language models (LLMs) with the ability to deeply analyze code repositories and commit histories. Unlike traditional coding tools, it can automatically trace the root causes of system crashes and generate repair patches by using semantic analysis, pattern recognition, and comprehensive processing of historical commit data. In the kBenchSyz benchmark tests targeting Linux kernel crashes, Code Researcher performed impressively, achieving a crash resolution rate of 58%, significantly surpassing SWE-agent's 37.5%. Additionally, Code Researcher was able to explore an average of 10 related files, compared to SWE-agent's 1.33 files, highlighting its powerful capability to deeply explore code repositories.

Broad Applicability and Practical Applications

Aside from the Linux kernel, Code Researcher has also demonstrated excellent versatility in testing open-source multimedia software. Through multi-faceted reasoning and global context collection, Code Researcher can provide high-quality crash repair solutions for various large codebases. This not only reduces the manual debugging burden on developers but also offers more efficient solutions for enterprise-level software maintenance. Microsoft AI stated that the launch of this tool will promote the automation process of system-level software development, saving developers valuable time.

Microsoft AI’s Code Researcher is not only a technical breakthrough but also a strong proof of the potential of AI in the field of software development. As AI agent technology continues to evolve, tools like Code Researcher are bringing us closer to artificial general intelligence (AGI). AIbase believes that the advent of this tool not only provides developers with powerful assistance but also sets a new benchmark for AI-driven development across the industry.

Paper: https://www.microsoft.com/en-us/research/publication/code-researcher-deep-research-agent-for-large-systems-code-and-commit-history/

相关资讯

OpenAI内斗时,Karpathy在录视频:《大型语言模型入门》上线

赶紧学习起来吧!OpenAI 的风波暂时告一段落,员工也忙着「干活了」。年初回归 OpenAI 的 Andrej Karpathy 最近做了一场关于大型语言模型(LLM)的 30 分钟入门讲座,但该讲座当时没录制。因此,他基于这场讲座重新录制了一个长达 1 小时的视频,希望让更多人看到和学习。视频的主题为《大型语言模型入门》,涵盖了 LLM 的推理、训练、微调以及新出现的 LLM 操作系统和 LLM 安全。视频主打「非技术性」,偏科普,所以更加容易理解。                               
11/24/2023 3:02:00 PM
机器之心

Meta开发System 2蒸馏技术,Llama 2对话模型任务准确率接近100%

研究者表示,如果 Sytem 2 蒸馏可以成为未来持续学习 AI 系统的重要特征,则可以进一步提升 System 2 表现不那么好的推理任务的性能。谈到大语言模型(LLM)的策略,一般来说有两种,一种是即时的 System 1(快速反应),另一种是 System 2(慢速思考)。其中 System 2 推理倾向于深思熟虑的思维,生成中间思维允许模型(或人类)进行推理和规划,以便成功完成任务或响应指令。在 System 2 推理中,需要付出努力的心理活动,尤其是在 System 1(更自动化思维)可能出错的情况下。因
7/15/2024 11:48:00 AM
机器之心

迈向System 2推理,100页论文硬核讲述Meta-CoT

Meta-CoT 通过显式建模生成特定思维链(CoT)所需的底层推理过程,扩展了传统的思维链方法。 「我们有一份关于『推理时间计算』的新研究,以及我们过去几个月一直在研究的内容! 我们提出了一些理论,说明为什么它是必要的,它是如何工作的,我们为什么需要它,以及它对超级智能意味着什么。
1/11/2025 3:41:00 PM
机器之心
  • 1