Reinforcement Learning for Code Optimization
The paper argues that timing-based rewards break naive RL for code speed, then reports large gains after redesigning the tests, rewards, and GRPO setup.
The authors introduce DMC-Optim, with larger optimization tests and a calibrated sandbox, to make execution time a usable signal. Their strongest configurations raise strict top-50% pass@1 from 18.0% to 31.3% on Qwen 2.5 7B and from 30.7% to 50.4% on CWM 32B. They say the gains hold at stricter speed percentiles while preserving pure-correctness scores. On LCB, CWM 32B wins up to 83% of median-sample speed comparisons against standard RLVR. HF Daily Papers' note
The authors introduce DMC-Optim, with larger optimization tests and a calibrated sandbox, to make execution time a usable signal. Their strongest configurations raise strict top-50% pass@1 from 18.0% to 31.3% on Qwen 2.5 7B and from 30.7% to 50.4% on CWM 32B. They say the gains hold at stricter speed percentiles while preserving pure-correctness scores. On LCB, CWM 32B wins up to 83% of median-sample speed comparisons against standard RLVR. HF Daily Papers' note
score 6