Yifan Zhou

Yifan Zhou · 周绎凡

CS + Math @ UCLA · Incoming MTS Intern @ OpenAI

I am a sophomore at UCLA studying Computer Science and Applied Mathematics. I’ve done some research in mech interp, using model diffing and activation patching to compare base models with their post-trained versions. My work studies how fine-tuning changes internal representations and how earlier computation shapes what later layers produce.

My broader interests also include multi-agent collaboration and alignment and efficient ML systems. I have studied how agent scaffolds shape collaboration and team performance, alongside agent behavior monitoring & evaluation and CPU-efficient inference systems (SimHash, int8 quantization, ONNX kernel fusion).

Previously, I interned as an Applied Scientist at Microsoft Research, 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.

Email GitHub Scholar LinkedIn X

Selected Publications

* equal contribution

A complete list of publications can be found on my Google Scholar page.

  1. ICML ICML
    component_contribution_poster.png
    Investigating Component Contributions in Multi-Agent ML Systems
    Junsung Kim*, Ilia Mireskandari*, Seungwan Son*, Yifan Zhou*, Khizer Shahid*, and Dylan Yihan Dai*
    ICML 2026 International Conference on Machine Learning (ICML)
    Multi-agent collaboration and alignment through systematic analysis of agent-scaffolding components.
    Paper Website
  2. Preprint
    pt-it-model-diff.png
    Same Targets, Different Computation: How Post-Training Divides Work Across Model Layers
    Yifan Zhou
    Preprint 2026
    Studies how post-training reorganizes dependencies between earlier and later computation, even when models learn the same target behavior.
    Paper Code

Industry Experience

OpenAI

Member of Technical Staff Intern (Incoming)

Jun 2026 – Sep 2026 · San Francisco, CA

Incoming Summer 2026; Integrity organization.

Microsoft Research

Applied Scientist Intern

Jan 2026 – Apr 2026 · Redmond, WA

Worked primarily on similarity-preserving hashing (SimHash) as an efficient approximation to cosine similarity, and also contributed to ONNX kernel-fusion optimizations. Mentored by Jinyu Li and Wenbin Zhu.

Celestra

Research Intern

Sep 2025 – Dec 2025 · Remote

Conducted research on multi-agent collaboration and alignment, co-authoring an ICML 2026 study of how agent-scaffolding components shape team performance across 4,000+ ablations. Also developed a modular multi-agent system achieving top performance on MLE-Bench.

Judgment Labs

Member of Technical Staff

Jan 2025 – Sep 2025 · San Francisco, CA

Built semantic search over long-form agent traces for trace bucketing and retrieval, and worked on RL-tuned LLM-as-a-judge for agent performance evaluation. Helped maintain judgeval.

Tencent

Student Researcher

Jun 2024 – Jul 2024 · Shenzhen, China

Reproduced an IBM Max-Cut QAOA paper on Tencent's superconducting quantum backend using TensorCircuit, then trained a small neural network to predict and mitigate per-circuit QAOA noise. Mentored by Yi-Chong Zheng.

Miscellaneous 🙃