CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generation

2026

Open publication workspace · Sign in to read the full PDF.

AI-generated summary

1. CUDA Agent is a large-scale agentic reinforcement learning system that generates high-performance CUDA kernels, significantly outperforming existing methods and compiler baselines.
2. This paper introduces CUDA Agent, a novel system for CUDA kernel generation that leverages a scalable data synthesis pipeline, a skill-augmented development environment, and advanced RL techniques for stable training.
3. CUDA Agent achieves state-of-the-art results on KernelBench, demonstrating superior performance over proprietary models and establishing LLM-based kernel generation as a competitive alternative to traditional compiler optimization.
Tags: CUDA, kernel generation, reinforcement learning, large language models, performance optimization

Check the original publication for accuracy and context.