Reproduce “Skill Blocks: How Should an Agent Load Its Skill? A Caching-Correct Comparison of Pre-load, On-Demand Tool-Loading, Progressive Disclosure, and Hybrid” 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
Hironobu Nakasuji
The paper compares four ways to expose reusable agent-skill procedures: preload all content, fetch blocks through a tool, select a reference, or summarize before fetching. In primary ScienceWorld tests, Skill Block reduced median effective input from 26,810 to 10,061 tokens, while completing 62 of 84 episodes versus Full’s 57. Paired tests detected no difference, and savings varied by task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.