Reproduce “How Much Skill Orchestration Can LLM Agents Compile Away? Execution-Grounded Specialization for Token-Efficient Skill Use” 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

Jiayi Ni, Zac Wang, Alex Zhang, Zheng Liu, Roopa Prabhu

SkillCache compiles a task trace into a reusable capsule, running mechanical steps while leaving judgment and recovery to the model. It uses execution evidence to bind values, and qualification checks structure, programs, and replay on inputs. Across 87 SkillsBench tasks, warm Sonnet runs used 247,000 tokens versus 1,313,000 cold, an 80.9% reduction; mean reward was similar in the tested setup.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.