An RL agent that beats Claude at writing CUDA kernels

What it does CUDA-Agent trains models via reinforcement learning to write optimized CUDA kernels. The project claims top results on KernelBench, outperforming the torch.compile baseline across difficulty levels and surpassing outputs from Claude Opus-4.6 and Gemini 3 Pro. The repository bundles a full agent workspace—complete with correctness checks and profiling tools—alongside a 6,000-sample training dataset named CUDA-Agent-Ops-6K.

The interesting bit Rather than simply scaling model size, the work applies agentic RL to the entire kernel development loop: an LLM implements kernels, compiles them, verifies correctness, profiles performance, and iterates. The training corpus is built by fusing operators from torch and transformers, then rigorously filtered to keep only executable, deterministic tasks that avoid overlap with standard benchmarks.

Key highlights

• Claims to be the first RL-trained model to outperform advanced proprietary models on high-performance CUDA generation.
• Achieves top results on KernelBench, with the strongest improvements on the hardest problems.
• Releases CUDA-Agent-Ops-6K, a curated dataset of 6,000 executable, deterministic operator fusion tasks.
• Ships a standardized agent_workdir with built-in correctness validation and performance comparison utilities.
• Includes SKILL.md, an expert-designed set of workflow constraints and optimization rules for the agent.

Verdict A solid starting point for researchers in ML-for-systems who want a reproducible agent loop for CUDA optimization. Probably overkill if you just need a drop-in kernel replacement without touching RL pipelines or agent workspaces.