ByteDance Seed and Tsinghua AIR Introduces CUDA Agent: A Large-Scale Agentic RL System for CUDA Kernel Generation

August 18, 2026 | Source: MarkTechPost

Tags: ByteDance, CUDA, reinforcement learning, GPU kernels, KernelBench, Seed1.6, agentic AI

ByteDance Seed and Tsinghua AIR's CUDA Agent trains an LLM via reinforcement learning to write GPU kernels that outperform torch.compile on 96.8% of KernelBench tasks at 2.11× geomean speedup — the first RL system to close the correctness-vs-speed gap in CUDA generation.

Details

GPU kernel optimization has long been a bottleneck: frontier LLMs write correct CUDA but slow CUDA, often underperforming torch.compile. CUDA Agent addresses this by placing Seed1.6 (23B active / 230B total parameters) inside a real CUDA development environment with profiling, correctness checks, and a sandboxed compiler, then training it with PPO for 150 steps at 131,072-token context.\n\nStarting from a base that outruns torch.compile on only 27.2% of tasks at 0.69× speedup, the trained agent achieves 98.8% pass rate and 96.8% faster-than-compile rate at 2.11× geomean — roughly 40 points ahead of Claude Opus 4.5 and Gemini 3 Pro on the hardest Level-3 split of KernelBench. The agent operates in a ReAct loop with Bash, Read/Write, Grep, and profiling tools — a setup mirroring OpenHands tooling.\n\nThe trained model weights are not released. But the research team makes public the CUDA-Agent-Ops-6K dataset (6,000 samples of synthesized fused-operator kernels), the SKILL.md specification, the milestone reward recipe, and anti-reward-hacking constraints. Teams at GPU cloud providers, inference companies, and latency-sensitive verticals can apply the methodology to open base models. Full replication requires 128 NVIDIA H20 GPUs.