Reinforcement Learning for Code Optimization

This paper studies reinforcement learning for code optimization, where correctness and execution speed must be optimized together. It introduces DMC-Optim, a large optimization benchmark with a calibrated sandbox, combines correctness and timing into the reward, and adapts GRPO and evaluation for sparse, noisy execution signals. Under strict top-50% pass@1, optimization-aware configurations improve Qwen 2.5 7B from 18.0% to 31.3% and CWM 32B from 30.7% to 50.4%. The reported gains are larger at stricter percentiles while preserving pure-correctness scores.