CUDA Agent Trains LLMs to Beat Compilers

ByteDance Seed and Tsinghua AIR released CUDA Agent, an agentic reinforcement learning system that trains a language model to write GPU kernels outperforming compilers. On KernelBench, Seed1.6 passes 74.0% of tasks but outruns torch.compile on only 27.2%, at 0.69x speedup.