Cathy Zhou is an AI engineer specializing in agent reinforcement fine-tuning: training reasoning models to navigate complex environments, use tools efficiently, and improve against measurable rewards. As a member of OpenAI’s fine-tuning team in 2025, she helped demonstrate how this approach can improve coding agents while reducing unreliable behavior and excessive inference costs.
At TreeHacks 2025, Zhou cocreated BrailleBot, an award-winning accessibility project combining document interpretation, image descriptions, and a modified 3D printer to produce braille. She also coauthored research on human-centered language models, contributing to its treatment of supervised fine-tuning.
Her work on tool-using agents centers on several practical engineering problems:
- Outcome-based reward design: Train coding agents against passing tests, relevant repository search, and verifiable results instead of rewarding polished but ineffective output.
- Isolated agent rollouts: Cognition’s Devin used separate virtual machines for individual training trajectories, preventing repository changes and shell commands from interfering across runs.
- Efficient tool use: Reinforcement fine-tuning can encourage parallel tool calls and eliminate unusually long trajectories, improving production latency and consistency.
- Reward-hacking defenses: GPU-kernel evaluations need correctness checks, static analysis, and actual performance measurements to catch nonexistent kernels and other misleading shortcuts.
Zhou’s public writing on agent training likewise emphasizes complex planning, multistep tool use, and custom rewards.
