跳转至

文章背景与核心概要

在当前的大语言模型应用中，当被要求分析实验时，智能体通常能够生成正常工作的代码，然而其分析结果是否具备科学上的可辩护性依然是一个巨大挑战。具备说服力的科学分析高度依赖于正确的程序化选择——例如选择领域公认的统计检验方法、权威的标识符命名空间以及必要的分析注意事项。为了解决这一痛点，本文作者推出了“科学智能体技能”（Scientific Agent Skills）开源库，涵盖了科学实践中16大领域的163项具体程序（包括基因组学、化学信息学、医学成像、研究设计与科学交流等）。

该论文的核心贡献在于提出了一种创新的技能组织形式：每项技能都表现为一个目录，其核心是一个经过版本控制、人类可读的指令文件，智能体仅在需要时才会动态加载该文件，同时目录中还包含参考资料和可执行脚本。研究对该文档语料库的属性进行了量化分析，发现所有163项技能的常驻描述仅占用200,000个Token窗口的7.1%，中位数工作流占用23.9%，但如果加载所有参考文件，部分工作流将面临上下文溢出的风险，这凸显了按需动态加载策略的必要性。

Scientific Agent Skills: A Library of Procedural Knowledge for Research Agents

Summary

When language-model agents are asked to analyze an experiment, they often generate working code, but whether that analysis is defensible remains a challenge. Defensible scientific analysis depends heavily on procedural choices—such as selecting field-accepted statistical tests, authoritative identifier namespaces, and necessary analytical caveats.

To address this, the authors introduce Scientific Agent Skills, an open library featuring 163 procedures across 16 domains of scientific practice (including genomics, cheminformatics, medical imaging, study design, and scientific communication). Each skill is structured as a directory built around a versioned, human-readable instruction file loaded dynamically by an agent only when required, alongside reference materials and runnable scripts.

Document Metadata

Abstract

大语言模型智能体在被要求分析实验时，通常会返回能够运行的代码。然而，该分析是否具备可辩护性则是另一个问题。可辩护的分析依赖于程序化选择：领域接受哪种检验、哪个标识符命名空间是权威的、以及结果必须伴随哪些注意事项。我们推出了“科学智能体技能”（Scientific Agent Skills），这是一个包含163个此类程序的开源库，横跨16个实践领域，包括基因组学、化学信息学、医学成像、研究设计和科学交流。每项技能都是一个目录，围绕一个带版本控制、人类可读的指令文件构建。智能体仅在任务需要时才加载该文件；该目录通常还包含参考材料和可执行脚本。我们未报告任务级别的评估和宿主选择率。我们对文档语料库的两个属性进行了测量：所有163项技能的常驻描述占200,000个Token窗口的7.1%，中位记录工作流占其中的23.9%，尽管如果加载每一个参考文件，46个工作流中有29个会溢出。该库采用开放许可，可在此处的链接获取。

A language-model agent asked to analyse an experiment will usually return working code. Whether the analysis is defensible is a different question. A defensible analysis depends on procedural choices: which test the field accepts, which identifier namespace is authoritative, and which caveats must accompany a result. We present Scientific Agent Skills, an open library of 163 such procedures in 16 areas of practice, including genomics, cheminformatics, medical imaging, study design and scientific communication. Each skill is a directory built around a versioned, human-readable instruction file. An agent loads the file only when a task calls for it; the directory often also contains reference material and runnable scripts. We report no task-level evaluation and no host selection rate. We measure two properties of the documentation corpus: the always-resident descriptions of all 163 skills cost 7.1% of a 200,000-token window, and the median documented workflow fits within 23.9% of it, although 29 of 46 would overflow if every reference file were loaded. Openly licensed and available at this https URL.

Corpus Findings & Resource Footprint

本研究评估了文档语料库的存储与上下文窗口开销： * 常驻成本： 所有163项技能的核心描述仅消耗标准200,000个Token上下文窗口的 7.1%。 * 工作流占用： 中位记录工作流适合在 23.9% 的Token窗口内运行。 * 溢出风险： 在评估的46个工作流中，如果有任何一个关联的参考文件被同时加载，29个 将超出上下文窗口限制，这凸显了按需动态加载策略的必要性。

The study evaluates the storage and context window footprint of the documentation corpus: * Always-Resident Cost: The core descriptions of all 163 skills consume 7.1% of a standard 200,000-token context window. * Workflow Footprint: The median documented workflow fits within 23.9% of the token window. * Overflow Risk: 29 out of 46 evaluated workflows would exceed the context window limit if every associated reference file were loaded simultaneously, emphasizing the necessity of on-demand, dynamic loading strategies.

Additional Resources & Full-Text Access

通过 arXiv 集成与开放获取出版标准进行维护。

Maintained via arXiv integration and open-access publishing standards.