Reproduce “Make Every Expansion Count: Learning to Search Efficiently in LLM Code Repair” 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
Elad Sharony, Ido Greenberg, Jialin Song, Shie Mannor, Gal Chechik, Eli Meirom
CodeTriage prioritizes repair edits before generating and testing revised programs. A small model uses test feedback, diagnoses, and edit descriptions to estimate how many expansions a successful path may need, then ranks pending edits. In offline replay of 5,001 held-out solvable Python repair trees, it needed 11.54 expansions to first fix versus 14.59 for AB-MCTS; this measures search order, not elapsed time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.