ByteDance, Tsinghua Unveil RL-Powered CUDA Agent

ByteDance’s Seed team and Tsinghua University’s Institute for AI Industry Research, known as AIR, have developed CUDA Agent to automate one of the most specialized parts of GPU computing. The system combines large language models with reinforcement learning to write, test and optimize CUDA kernels, work that typically requires experienced engineers and extensive manual tuning. Better kernel generation could reduce development costs and improve the efficiency of AI training and other high-performance workloads.

The researchers said CUDA Agent operates at scale in real sandbox environments, using execution feedback and Proximal Policy Optimization, or PPO, to refine its code-generation strategy. In KernelBench testing, the system delivered stronger acceleration results than several leading frontier models and produced kernels that, in some cases, outperformed compiler-generated implementations. The report did not specify a release date or disclose the complete benchmark scores, limiting direct comparisons of absolute speed gains across hardware and workloads.