RLPF Trains Code Agents for Faster, Correct Code Generation

Current methods for training AI code generation models often stop at ensuring functional correctness, overlooking the critical aspect of runtime efficiency, especially for systems code. Two programs might pass the same tests but differ significantly in speed. The challenge lies in providing meaningful feedback for efficiency, as runtime is only relevant for correct programs and offers little guidance when programs fail to compile or execute. To address this, a new technique called RLPF (Reinforcement Learning from Performance Feedback) has been developed. RLPF transforms execution outcomes into a staged reward system. For programs that fail, feedback is provided based on their execution progress, offering guidance even before correctness is achieved. For correct programs, the reward is based on their relative efficiency improvement compared to a baseline or expert reference. Fine-tuning Qwen3-32B with RLPF on PerfCodeBench dramatically increased correct-and-runnable solutions and improved relative efficiency. The trained model became competitive with stronger open-weight systems and showed some transferability of optimization behavior. This demonstrates that RLPF enables code agents to not only pass tests but also to actively optimize for performance.