CUDA Agent: Agentic RL Outperforms Opus 4.5 on GPU Kernel Generation

An agentic RL system that autonomously generates optimized CUDA kernels, outperforming torch.compile by 100% on KernelBench Level-1/2 and beating Claude Opus 4.5 and Gemini 3 Pro by ~40% on the hardest tasks. Demonstrates that agentic RL can surpass both compiler-based and frontier-LLM approaches. arXiv 2602.24286