ByteDance Seed and Tsinghua AIR have introduced CUDA Agent, an agentic reinforcement learning system that trains the Seed1.6 large language model to write GPU kernels that outperform compilers.

  • The system uses PPO training over 150 steps with a 131,072-token context within a profiling sandbox and correctness checks.
  • It achieves a 98.8% pass rate and a 96.8% faster-than-torch.compile rate on the KernelBench benchmark, yielding a 2.11× geometric-mean speedup.
  • The research team released the CUDA-Agent-Ops-6K dataset, SKILL.md spec, and reward recipes, but not the trained agent weights.
  • Ablations show that removing the agent loop drops the faster-than-compile rate from 96.8% to 14.1%, highlighting the importance of the agentic approach.

The work demonstrates that placing an LLM in a real CUDA development environment with iterative profiling can significantly close the performance gap between generated and hand-written or compiler-generated kernels.