ByteDance and Tsinghua unveil CUDA Agent—AI system uses reinforcement learning to auto-generate GPU code; achieves 2.11x speedup over PyTorch, 98.8% success rate on 6,000 synthesized tasks

CUDA Agent | Large-Scale Agentic RL for CUDA Kernel Generation

CUDA Agent High-Quality Training Tasks via a Scalable Data Pipeline CUDA Agent is a large-scale agentic reinforcement learning system that develops robust CUDA kernel optimization ability through scalable data synthesis, a skill-augmented execution environment, and stable long-horizon RL training. Hanlin Wu1,2,3*, Qiying Yu1,2,3, Huan-ang Gao1,2,3, Jiahao Li1, Chengquan Jiang1, Weiqiang Lou1, Yufan Song1, Hongli Yu1,2,3, Jiaze Chen1,3, Wei-Ying Ma2,3, Ya-Qin Zhang2,3, Jingjing Liu2,3, Mingxuan W...