Jiayi Weng 翁家翌

trinkle23897 [at] gmail [dot] com

I am a research engineer at OpenAI, where I built and optimized reinforcement learning systems to align foundation models. My work spans high‑throughput RLHF pipelines and tooling that lets researchers iterate easy and fast.

Previously, I received my bachelor's degree from Tsinghua University and my master's degree from Carnegie Mellon University. I was a research engineer at Sea AI Lab in Singapore, advised by Min Lin from May, 2021 to September, 2021. I spent a wonderful time at TSAIL, working with Professor Hang Su and Jun Zhu in the field of Reinforcement Learning from March, 2018 to June, 2020.

I created Tianshou (PyTorch RL library, ) and EnvPool (ultrafast RL environment executor, ). I enjoy open‑source and systems that make RL research easier and faster.

CV  /  GitHub  /  Google Scholar  /  LinkedIn  /  Ins  /  Zhihu

Since July 2022, I have been a research engineer at OpenAI’s Post‑Training team, designing and scaling RLHF pipelines. My contribution to OpenAI's published work includes:

• ChatGPT – initial release effort (6th author)
• GPT‑4 – RL infra author
• GPT‑4V – multimodal RL
• GPT‑4 Turbo – post‑training
• GPT‑4o – post‑training infra lead
• o‑series models – post‑training infra
• Operator – early RL effort
• GPT‑4.5 – post‑training
• RFT (reinforcement fine-tuning) – core RL infra author

For the summer of 2021, I was a full-time research engineer at Sea AI Lab in Singapore, where I worked with Min Lin and Professor Shuicheng Yan. We created a ultrafast RL vectorized environment EnvPool that can train Atari Pong and Mujoco Ant in 5 mins on a laptop, and has released on GitHub.

For the summer of 2019, I was a visiting student researcher at the Montreal Institute for Learning Algorithms (MILA), where I worked with Min Lin and Professor Yoshua Bengio on the Consciousness Prior based on Transformer architecture. [Photo]

Prior to that, I worked on the reinforcement learning algorithm based on the VizDoom platform with Professor Hang Su and Jun Zhu in TSAIL. We proposed an environment-aware hierarchical reinforcement learning architecture and achieved first place in VizDoom AI Competition 2018 Single Player Track(1). I am also the main contributor of Tianshou, an elegant, flexible, and superfast PyTorch deep Reinforcement Learning library.

I'm also engaged (but an amateur) in Photography [1] [2] / Computer Graphics / Web Security.

, Credits