Recoleta Item Note
CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generation
This paper proposes CUDA Agent, a large-scale agentic reinforcement learning system for CUDA kernel generation, designed to help models genuinely learn GPU kernel optimization rather than merely perform shallow…
cuda-kernel-generationagentic-rlcode-optimizationgpu-programmingcode-intelligence
Summary
This paper proposes CUDA Agent, a large-scale agentic reinforcement learning system for CUDA kernel generation, designed to help models genuinely learn GPU kernel optimization rather than merely perform shallow iterative patching. It targets the high-barrier scenario of high-performance code generation and reports new state-of-the-art results on KernelBench.
Problem
- CUDA kernel optimization is crucial for modern deep learning, but it depends heavily on hardware expert experience and is difficult to automate and scale.
- Although existing LLM methods can write general-purpose code, they still lag behind compiler-style systems in CUDA kernel generation, indicating that the models lack true low-level optimization ability.
- Existing methods mainly rely on training-free refinement or fixed multi-turn execution-feedback fine-tuning. These mechanisms struggle to fundamentally improve the model's CUDA optimization ability, so the performance gains are limited.
Approach
- The core idea is to turn CUDA kernel development into an agentic reinforcement learning task with verifiable, measurable rewards, allowing the model to learn in a closed loop of "write code - verify - analyze performance - improve again."
- The authors build a scalable data synthesis pipeline to generate enough CUDA training tasks and trajectories to support large-scale RL training.
- The system provides a skill-augmented CUDA development environment, including automated verification and profiling, thereby producing more reliable reward signals rather than relying only on textual preferences or vague feedback.
- The paper also introduces RL algorithmic techniques needed for stable training, supporting large-scale agentic learning and the gradual development of CUDA kernel optimization expertise.
Results
- On KernelBench, CUDA Agent achieves faster rate improvements over the baseline system mentioned in the paper: Level-1 +100%, Level-2 +100%, and Level-3 +92%.
- On the hardest KernelBench Level-3 setting, CUDA Agent outperforms the strongest proprietary models, Claude Opus 4.5 and Gemini 3 Pro, by about 40%.
- The paper claims state-of-the-art results on KernelBench, suggesting that agentic RL does not merely improve formatting or executability, but significantly enhances CUDA optimization ability.
- The abstract does not provide more fine-grained absolute scores, variance, training scale, or ablation numbers, so the currently confirmable quantitative results are mainly the faster-rate relative improvements above.
Link
Built with Recoleta
Run your own research radar
Turn arXiv, Hacker News, OpenReview, Hugging Face Daily Papers, and RSS into local Markdown, Obsidian notes, Telegram digests, and a public site.