Reproduce “Learning Meta-Skills for Agent Harness Design in Test-Time AI4AI” 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
Cheng Qian, Kunlun Zhu, Beibin Li, Zhenhailong Wang, Heng Ji
The AI Builder learns reusable meta-skills from development executions, then uses them to construct task-specific harnesses for a fixed AI Target. It supplies instructions, resources, and executable support without changing model weights. Across six model-benchmark settings, full-bank meta-skills averaged 65.31%, 8.95 pp above no-skill construction and 12.02 pp above direct delivery; these gains measure Target performance under fixed execution limits.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.