News Center

location: Home > News Center > News > Content

AIR Research | Breaking the CUDA Code Generation Bottleneck for LLMs: An Agentic Reinforcement Learning System

Source:       Time:2026-03-04

GPU kernel optimisation is the last mile of efficiency in modern deep learning infrastructure, yet it has long been the preserve of a small group of engineers with deep hardware expertise. AIR and ByteDance Seed have jointly developed CUDA Agent, which applies large-scale agentic reinforcement learning to high-performance CUDA kernel generation. Project page: https://cuda-agent.github.io/

https://air.tsinghua.edu.cn/info/1007/2523.htm


上一条:AIR Research | Hyper Diffusion Planner: Putting Diffusion Models on Real Roads with a 10× Gain in End-to-End Closed-Loop Driving Performance 下一条:Cell | Tsinghua AIR Releases PocketXMol, an Atom-Level Generative AI Model for Drug Discovery

【关闭】

Relevant news

Email:Airoffice@air.tsinghua.edu.cn
Tel:(010)82151160

Address:12 / F, block C, Qidi science and technology building, Tsinghua Science and Technology Park, Haidian District, Beijing

wechat

Jing ICP Bei No. 15006448 | all rights reserved@ Institute of intelligent industry, Tsinghua University