←新闻详情

InternLM 原创项目更新于 10月14日 12:28

InternLM 开源 Skill-as-Pseudocode:把 markdown 技能库重构为类型化伪代码契约

GitHub - InternLM/Skill-as-Pseudocode: [EMNLP 2026 Findings] Skill-as-Pseudocode: refactoring markdown skill libraries into typed pseudocode contracts for LLM agents. · GitHub

AI 导读

InternLM 发布 SaP(Skill-as-Pseudocode)参考实现,将自由形式的 markdown 技能库转换为类型化伪代码契约,用于 LLM 智能体。该流水线在 ALFWorld 和 SkillsBench 上对照 Graph-of-Skills 检索基线进行评测,在 skills_500 上以 gpt-4o-mini、temperature 0 运行的总 LLM 成本约 1.30 美元。仓库采用 MIT 许可,并附带可直接跳过流水线的预构建 artifacts。

正文

当前显示原文,中文全文译文尚未就绪。

InternLM; Skill-as-Pseudocode

Repository navigation

NotificationsYou must be signed in to change notification settings; Fork0 (0); Star4 (4)You must be signed in to star a repository

[EMNLP 2026 Findings] Skill-as-Pseudocode: refactoring markdown skill libraries into typed pseudocode contracts for LLM agents.

Repository files navigation

README; MIT license

This repository contains the reference implementation of SaP — a verified pipeline that converts free-form markdown skill libraries into typed pseudocode contracts, plus the evaluation hooks used to measure SaP against the Graph-of-Skills retrieval baseline on ALFWorld and SkillsBench.

SaP as verified refactoring. Repeated prose spans in parent skills (A) feed a numbered, checked pipeline (B): form a candidate cluster, draft a child contract κ, verify it, bind concrete arguments and clean residual conflicts, then emit an agent-facing pseudocode skill. The resulting representation (C) separates what the skill does (typed signature), what local arguments this parent supplies (grounded invoke), and how the environment is actually called (concrete action template).

If you use this code, please cite our paper.

Repository layout

Quick start

To reproduce the pipeline from scratch (replace the artifacts):

Total LLM cost for the full pipeline on skills_500 is approximately $1.30 against gpt-4o-mini at temperature 0.

Pre-built artifacts

For the impatient, artifacts/ ships everything the agent needs at run time so you can skip the pipeline entirely:

Environment

The pipeline uses a handful of environment variables (with sensible defaults) so it can be relocated without code edits:

Citation

If you find this work useful, please cite our paper:

[EMNLP 2026 Findings] Skill-as-Pseudocode: refactoring markdown skill libraries into typed pseudocode contracts for LLM agents.

来源

InternLM 原创项目