Reproduce “Harness Self-Evolution for Long-Horizon Terminal 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
Ruhan Wang, Yucheng Shi, Zongxia Li, Junyao Yang, Zhongzhi Li, Yue Yu, Haitao Mi, Dongruo Zhou and 1 more
RiseHarness improves terminal-agent harnesses without changing evaluated model weights. It compares mixed outcomes using agent and harness traces plus a behavior-to-code Handbook, adapts designs, and retains changes after checks and repeated evaluation. On Terminal-Bench Hard with DeepSeek-V4-Flash, Pass@1 rose from 56.75% to 64%, and four-run Pass₄ rose from 37% to 51%; these comparisons suggest harness issues but do not establish causation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.