CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generation (arxiv.org)

A significant advancement in GPU kernel optimization was announced with the release of CUDA Agent, a large-scale reinforcement learning system designed to enhance CUDA kernel generation. Traditional methods for optimizing GPU performance require specialized knowledge, leading to inefficiencies in producing high-quality CUDA code. CUDA Agent addresses this issue by integrating a scalable data synthesis pipeline and an augmented development environment that automates verification and profiling, resulting in reliable reward signals for training. This innovative approach allows for stable reinforcement learning that significantly enhances the model's CUDA optimization capabilities. CUDA Agent has demonstrated impressive performance improvements on the KernelBench benchmarks, achieving speeds that are 100%, 100%, and 92% faster on Level-1, Level-2, and Level-3 splits, respectively. Notably, it outperforms leading proprietary models like Claude Opus 4.5 and Gemini 3 Pro by approximately 40% on the challenging Level-3 setting. This breakthrough not only sets a new standard for CUDA kernel generation but also underscores the transformative potential of reinforcement learning in optimizing complex programming tasks, marking a substantial leap forward for the AI/ML community in high-performance computing.