I train large language models to act, not only to answer.

AI Research Scientist at WRITER, working on reinforcement learning and post-training for agents. I also go by George.

Zhizhuo Yang holding a camera in Yosemite Valley, with El Capitan behind him

I'm Zhizhuo Yang (杨智卓), a machine learning researcher interested in how agents perceive, act and learn. Today those agents are language models. At WRITER I own LLM post-training end to end, as the primary builder of the agentic reinforcement learning stack: data synthesis, verification and rewards, and training a 744-billion-parameter mixture-of-experts model with the production agent harness as its environment. I also built the synthetic tool-use data pipeline behind Palmyra x6, WRITER's shipped agentic model.

I came to machine learning through optics and virtual reality, with degrees from Beijing Jiaotong University and Beijing Institute of Technology. For my Ph.D. at Rochester Institute of Technology, advised by Dr. Alexander G. Ororbia II and Dr. Reynold Bailey, I started with deep learning for eye tracking and for how the eyes and head move together, and from 2020 I worked on active inference, which became my dissertation on perception, action and learning. Internships at Facebook and Meta took me into rendering for VR headsets and sensor fusion for glove tracking.

I have mentored students since my master's, and at WRITER I supervised an intern working on diffusion language models. Away from work I love music and movies, play basketball and badminton, travel, and hike with a camera. I also play a little guitar.

The way here

From optics and virtual reality, through eye tracking and active inference, to training language models that act.

  1. 2010Optical information scienceB.S., Beijing Jiaotong University
  2. 2014Virtual and augmented realityM.S., Beijing Institute of Technology
  3. 2017Eye tracking and gazePh.D. research, Rochester Institute of Technology
  4. 2019Rendering for VR headsetsInternships at Facebook, 2019 and 2020
  5. 2020Active inferencePh.D. dissertation, completed in 2024
  6. 2023Sensor fusion for glove trackingMeta, internship and student researcher
  7. 2025Reinforcement learning for LLM agentsAI Research Scientist, WRITER

Selected work

All research

WRITER, 2025 to 2026

Reinforcement learning for a 744B-parameter agent

End-to-end GRPO training across 8-node GPU clusters, with the production agent as the environment.

Asynchronous one-step off-policy training cut per-step non-generation time from 170 seconds to 24 seconds.

The training loop: synthetic tasks go to the production agent harness, where the 744B mixture-of-experts policy acts; verifiers score the result as rewards, and a GRPO update improves the policy. Synthetic tasks Policy 744B mixture of experts Agent harness the production one Verifiers and rewards GRPO update on the Slime framework acts trajectories reward

Technical report, arXiv 2026

Palmyra x6: an agentic, tool-use model

WRITER's shipped agentic model, post-trained on data from a synthetic tool-use pipeline I built. It has the top BFCL Core score (0.785) among the compared models.

Grouped bar chart of scores on six agentic benchmarks, FinanceBench, IFBench, Finance Agent, FinanceQA, MCP-Atlas and BFCL Core, with each model's mean in a shaded group at the right. Palmyra x6 is the first, blue bar in each group. It has the tallest bar on BFCL Core, at 0.785, and the tallest bar in the mean group.

Recent news

  • Our Palmyra x6 technical report is on arXiv.

  • Today, I started my first full-time job as an AI researcher at WRITER! I’ll be working on LLMs post-training, AI agents, and many more exciting topics in the generative AI field.

  • Super excited to announce that our paper SR-AIF: Solving Sparse-Reward Robotic Tasks from Pixels with Active Inference and World Models has been accepted at ICRA 2025!

  • Thrilled to announce that I successfully defended my PhD dissertation today and officially became Dr. Yang!

Outside work

All photographs

I photograph landscapes on my hikes, from Sequoia to the Olympic Peninsula.