ByteDance Seed and Tsinghua AIR Introduces CUDA Agent: A Large-Scale Agentic RL System for CUDA Kernel Generation

CUDA Agent is a reinforcement learning system that trains a large language model to write GPU code (CUDA kernels) that runs faster than compiler-generated code. Frontier AI models already produce correct GPU code but often generate slower versions, achieving only a 0.69× speedup compared to standard compilers. By placing the model in a real development environment with profiling tools and correctness checks, then training it with reinforcement learning for 150 steps, CUDA Agent improved performance to a 2.11× speedup and a 96.8% rate of outperforming the standard compiler on benchmark tasks. This matters for AI infrastructure, GPU computing, and any application where GPU kernel performance directly affects latency and cost, though the trained agent itself is not publicly released and full replication requires substantial computational resources.