Reproduce “ExpertCrowd-SWE: Measuring Frontier Coding Agents on Expert-Authored Engineering Tasks in Unfamiliar Codebases” 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
Josip Vrdoljak, Akash Singh, Bas Steunebrink, Alex Yarotsky, John Lorenzo, Achille Nicoletti, Sam Pigott
EC-SWE evaluates whether autonomous agents can make maintainer-acceptable changes from short requests in unfamiliar repositories. Practicing engineers write unpublished, multi-file tasks, graded with rubrics that award partial credit and enforce blockers and rejection gates. Across 100 tasks, Claude Opus 5.5 averaged 67.3% and GPT-6 Astra 63.7%, but their difference was not statistically established; scores reflect this benchmark, not model quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.