Reproduce “SkillRAE: Agent Skill-Based Context Compilation for Retrieval-Augmented Execution” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiangcheng Meng, Shu Wang, Yixiang Fang
SkillRAE is a Retrieval-Augmented Execution (RAE) framework designed to enhance LLM-based agent performance by compiling retrieved skills and their subunits into a compact, task-specific context. The method, developed at The Chinese University of Hong Kong, Shenzhen, achieves a 29.26% mean reward on SkillsBench and an 84.59% aggregate score on AgentSkillOS, outperforming existing baselines by explicitly organizing procedural knowledge for execution.
Key insights
- Explicitly compiling retrieved skills and their fine-grained subunits into a structured, task-specific context significantly enhances LLM agent performance in skill execution.
- A multi-level skill graph, encompassing skill communities, skills, and subunits, provides crucial evidence sources for both effective skill retrieval and intelligent context generation.
- The 'context compilation' stage is a critical missing link in existing RAE frameworks, transforming raw retrieved information into actionable instructions for agents.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.