Stable Asynchrony: Variance-Controlled Off-Policy RL for LLMs
Luke J. Huang, Zhuoyang Zhang, Qinghao Hu, Shang Yang, Song Han
Luke J. Huang
lukh23@mit.edu Physics + CS @ MIT
Hey! I'm Luke. I'm a Physics + CS student at MIT and a Member of Technical Staff Resident at OpenAI. My work focuses on reasoning, efficient AI computing, reinforcement learning for LLMs, and fast generative models.
Publications & Preprints
Locality-Aware Parallel Decoding for Efficient Autoregressive Image Generation
Zhuoyang Zhang*, Luke J. Huang*, Chengyue Wu, Shang Yang, Kelly Peng, Yao Lu, Song Han
ForeAct: Steering Your VLA with Efficient Visual Foresight Planning
Zhuoyang Zhang, Shang Yang, Qinghao Hu, Luke J. Huang, Jianing Hou, Yifu Sun, Yao Lu, Song Han
AI-Driven Robotics for Free-Space Optics
Shiekh Zia Uddin*, Sachin Vaidya*, Sarthak Choudhary, Zhuo Chen, Ronald K. Salib, Luke J. Huang, Dirk R. Englund, Marin A. Soljacic
Experience
OpenAI (May 2026 - Present)
Member of Technical Staff Resident, San Francisco
Working on pretraining.
Applied Compute (Jun 2025 - Sep 2025)
Research Scientist Intern, San Francisco
Built large-scale RL training infrastructure for out-of-distribution reasoning tasks, including an asynchronous RL framework and a Torchtitan/FSDP backend for 100B+ MoE models.
Han AI Lab (Aug 2024 - Present)
Researcher, Cambridge
Researching robust asynchronous RL, efficient autoregressive image generation, and low-latency generative models for VLA systems.
Education
Massachusetts Institute of Technology (Fall 2024 - Present)
Physics + CS, 5.0/5.0 GPA
Selected Coursework
Distributed Systems, Mathematical Statistics, Deep Learning, Information Theory, Inference and Information, Design and Analysis of Algorithms, Algebra I, Quantum Physics III.
Honors & Awards
Selected
Putnam Top 200, 2024 U.S. IPhO Team and Gold Medalist, Regeneron STS Finalist, Math Olympiad Program attendee, USA(J)MO winner, and RSI Scholar.