Reproduce “Optimize the Harness, Not the Model: Fitting a Minimal Agent to a Self-Hostable LLM for Investigative Analysis” 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
Shai Nahum Gefen, Vered Chassida, Dmitry Basin, Guy Feigenblat
CLEO builds a small harness around a fixed model, adding tools or control changes only when evaluation shows gains and review finds them general, auditable, and lean. On held-out intelligence questions, GLM-5.1 with CLEO reaches 83.8%, near Claude Code SDK’s 85.0%, with roughly 3,000 versus an estimated 512,000 harness lines. The single judge was not systematically validated; the SDK size is unverified.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.