Yifan Zhou

I am a student at University of California, Los Angeles studying Computer Science and Applied Mathematics. I am currently conducting independent research in mechanistic interpretability, developing methods to analyze how post-training reshapes model behavior — using first-divergence model diffing to pinpoint the earliest token where a pretrained model and its instruction-tuned descendant disagree, then decomposing how upstream state and late-stack computation jointly produce that disagreement.

My broader interests also include reinforcement learning, multi-agent systems, and efficient ML systems. I have worked on multi-agent component analysis, agent behavior monitoring & evaluation, and CPU-efficient inference systems (SimHash, int8 quantization, ONNX kernel fusion).

Previously, I interned as an Applied Scientist at Microsoft, working on SimHash and efficient ML systems. I have also worked at Celestra, Judgment Labs, and Tencent. I will be joining OpenAI (Integrity) as a Member of Technical Staff Intern in Summer 2026.