Reproduce “Meta$^2$: Precomputing Reusable Knowledge for Online Harness Evolution in Long-Horizon Agents” 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
Lingzhi Yuan, Yiguang Cai, Weize Liu, Siqi Yang, Feijiang Han, Yongkyun Lee, Udari Madhushani Sehwag, Furong Huang
Meta² teaches an evolver reusable guidance for changing a frozen model’s task harness during a run, rather than selecting one harness for every task. Its Meta-Skill guides when to edit, hold, or avoid changes; edits remain run-local, without held-out verifier feedback. With GPT-5.6-Terra, it reports mean@5 81.00% on Terminal-Bench 2.1 and mean@3 44.44% on DeepSWE 1.1, on two tested benchmarks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.