Reproduce “Correctness vs Efficiency: Behavior Control with Weight-Space Extrapolation in Code RL” 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

Kunhao Zheng, Pierre Chambon, Juliette Decugis, Jonas Gehring, Taco Cohen, Benjamin Negrevergne, Gabriel Synnaeve

The authors train models from a shared start with loose and strict test-coverage rewards, then move along their weight difference. In 32B tool-use tests, extrapolation extends the trade-off between resource-limit and correctness failures while solve rate stays flat from α = −0.5 to 1.5. Extrapolated code runs 1.14× faster on 141 problems both models solved; this does not establish a general reward-hacking fix.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.