ByteDance and Tsinghua AIR release CUDA Agent for GPU kernel generation

ByteDance Seed and Tsinghua AIR have released CUDA Agent, an agentic reinforcement learning system designed to generate highly optimized GPU kernels. Built on the proprietary 230B-parameter Seed1.6 model, the system operates in a secure CUDA sandbox and is trained using PPO. On the KernelBench benchmark, CUDA Agent achieved a 98.8% pass rate and outperformed torch.compile on 96.8% of tasks, delivering a 2.11× geometric-mean speedup. While model weights remain closed, the team has open-sourced the 6,000-sample CUDA-Agent-Ops-6K dataset.

CUDA Agent system architecture

• ▪CUDA Agent operates inside a real CUDA development environment featuring profiling, correctness checks, and a permission-locked sandbox.
• ▪The profiling sandbox used to develop CUDA Agent required 128 NVIDIA H20 GPUs.
• ▪CUDA Agent is built on Seed1.6, a proprietary Mixture-of-Experts model with 23 billion active and 230 billion total parameters, whose weights are not publicly released.
• ▪ByteDance Seed and Tsinghua AIR released CUDA Agent, an agentic reinforcement learning system that trains a large language model to write GPU kernels.

KernelBench performance results

• ▪The base model Seed1.6 passes 74.0% of KernelBench tasks but outperforms torch.compile on only 27.2% of them, resulting in a 0.69× geometric-mean speedup.
• ▪On the KernelBench benchmark, CUDA Agent achieved a 98.8% pass rate and a 96.8% faster-than-torch.compile rate, yielding a 2.11× geometric-mean speedup over compile.
• ▪On the Level-2 split of KernelBench, CUDA Agent achieved a 100% pass rate and a 2.80× geometric-mean speedup over torch.compile.
• ▪On the Level-3 split of KernelBench, CUDA Agent achieved a 94.0% pass rate and a 90.0% faster rate, outperforming Claude Opus 4.5 by 40 percentage points.

CUDA-Agent-Ops-6K dataset synthesis

• ▪The research team synthesized the CUDA-Agent-Ops-6K dataset, which contains 6,000 samples, by crawling reference operators from the torch and transformers libraries.
• ▪The CUDA-Agent-Ops-6K dataset, the SKILL.md specification, and the reward and warm-up recipes have been released publicly.
• ▪The CUDA-Agent-Ops-6K dataset consists of 83.77% two-operator compositions, excluding samples with an AST similarity above 0.9 to any KernelBench task.

Reinforcement learning training methodology

• ▪The agent loop of CUDA Agent mirrors OpenHands tooling, utilizing a ReAct pattern with tools such as Bash, Read/Write, Edit/MultiEdit, and NotebookEdit.
• ▪CUDA Agent was trained using Proximal Policy Optimization for 150 steps at a 131,072-token context window.
• ▪Ablation studies showed that removing the agent loop from CUDA Agent dropped its faster rate versus torch.compile from 96.8% to 14.1%.

Discrete milestone reward design

• ▪CUDA Agent uses a discrete milestone reward system with values of -1, 1, 2, or 3, rather than a raw speedup ratio.
• ▪The discrete milestone reward system used in CUDA Agent outperformed a raw speedup reward by 36.4 percentage points on the faster-than-compile rate.

Anti-hacking verification measures

• ▪To prevent reward hacking, CUDA Agent implements five countermeasures, including permission-locked verification scripts and context managers that forbid torch.nn.functional fallbacks.
• ▪CUDA Agent's anti-hacking measures include testing against five random inputs, profiling with device synchronization and warm-up, and omitting a web search tool.

1 source