Meta FAIR's RL Method Nearly Doubles Code-Optimization Pass Rates

Most reinforcement-learning work on code models optimizes for one thing, does the program pass the tests. A new paper from Meta FAIR tries a harder question: whether RL can also push a model to write faster code without breaking correctness, and what actually stops that from working when you naively add execution time to the reward signal.

The team built a training and evaluation setup called DMC-Optim, with a calibrated sandbox and an adapted version of the GRPO algorithm tuned for what they describe as the sparser, noisier timed-execution setting. The numbers they report are meaningful. Strict top-50% pass@1 on Qwen 2.5 7B rose from 18.0% to 31.3%, and on their own CWM 32B model from 30.7% to 50.4%. On the LCB benchmark, CWM 32B is said to win up to 83% of median-sample speed comparisons against a standard RLVR baseline.

The more interesting technical point is that once timing drives the reward, all the usual RL failure modes get louder. Measurement noise, reward sparsity, and GRPO instability start to swamp the signal, and much of the paper is about the tricks needed to keep training stable. Their robust variant reportedly delivered 100% to 200% improvement over standard RLVR when the sandbox timing was deliberately degraded, which is the kind of stress test that matters if any of this is going to run outside a research cluster.

The honest caveat sits inside their own numbers. The model's complexity-class improvement rate is 14%, compared with 28% for humans on the same problems. In plain terms, the system is getting better at squeezing constants out of a given algorithm rather than picking a better algorithm in the first place, which is the part of optimization that tends to matter most on real workloads. What the reporting doesn't give you is training compute cost, or any read on how well the calibrated sandbox generalizes to codebases that look nothing like competitive-programming problems.

If the method survives outside its own benchmark, the shape of the opportunity is clear enough. Post-training recipes that push models past 'passes the tests' into 'runs fast on the tests' are a missing piece for LLMs sitting inside CI, and Meta FAIR publishing a template lowers the bar for smaller labs to try the same thing on their own models.