RLPF lifts runnable code from 11.1% to 54.6%

RLPF fine-tuning raised Qwen3-32B’s correct-and-runnable rate on PerfCodeBench from 11.1% to 54.6% and its relative efficiency from 8.1% to 38.6%, according to the new reinforcement-learning study. The method ranks failed programs by execution progress, then ranks correct programs by runtime improvement toward an expert reference, supplying a useful reward before and after correctness. For teams building the verified loops described in the coding-agent stopping-rule debate, the operator implication is narrow but important: passing tests should be the gate, not the final reward, when generated systems code has a performance budget.