Reproduce “Recursive Self-Improvement on a Budget: Scoring Code Edits Before They Run” 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
Erfan Darzi, Sree Bhargavi Balija, Sheng He
The paper introduces a linear readout to choose which scaffold edit a coding agent should test next. It combines the parent’s solve rate with eight coordinates representing the diff, fits the readout on earlier edits, and freezes it during search. On held-out HumanEval+ with GPT-4o, it achieves 87.0% versus 84.2% for parent-rate selection; evidence is limited to this tested setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.