Fine-Tuning GPT-5 for GPU Kernel Generation

Researchers have fine-tuned GPT-5 using reinforcement learning to generate highly efficient GPU kernels, a notoriously complex task. The model, called Makora, achieved a 33.3 percentage point jump in correctness (from 43.7% to 77.0%) and now outperforms the standard PyTorch TorchInductor compiler on 72.9% of problems, delivering a geometric mean speedup of 2.12x. This demonstrates RL can unlock LLM capabilities in data-scarce, specialized domains where traditional supervised fine-tuning fails.