Meta paper flags reinforcement learning hurdles in code optimization, introduces DMC-Optim benchmark

Meta AI’s FAIR team published a paper on July 29 saying reinforcement learning faces major obstacles in code speed optimization, including heavy timing-measurement noise and sparse rewards, which make standard algorithms unstable. The team introduced a benchmark called DMC-Optim to study the problem in a more controlled setting. The framework rebuilds the feedback system used in AI model training and combines correctness rewards with execution-speed rewards inside an offline simulator, creating a calibrated sandbox where timing noise can be adjusted. Results cited in the paper show the benchmark lifted model pass rates by as much as 64%, while performance gains under amplified noise conditions reached 100% to 200%. In the reported tests, Qwen 2.5 7B improved its pass rate from 18.0% to 31.3%, and CWM 32B rose from 30.7% to 50.4%. The paper also said CWM 32B outperformed traditional reinforcement learning methods 83% of the time on the LCB benchmark. The development was reported by CryptoBriefing.