Multi-turn RL with Structural and Performance Aware Rewards for CUDA Kernel Generation

CudaPerf addresses the problem that existing RLVR methods for CUDA kernel generation overlook structural code properties like memory coalescing and occupancy, relying only on outcome-based signals. The method proposes a two-stage reflective RL framework with offline pairwise ranking and online RL training that jointly optimizes correctness, performance, and structural efficiency via a unified reward. Experimental evidence shows CudaPerf achieves up to 5X and 3.32X speedup improvements and 17% and 7% correctness gains over Qwen-3-32B and CUDA Agent on C-to-CUDA and PyTorch-to-CUDA benchmarks. This matters because it demonstrates that incorporating structural code-aware rewards significantly enhances LLM-driven optimized code generation for high-performance computing.