Yifan Zhou

Yifan Zhou

Software Engineer at Google

Hello! I’m Yifan Zhou, a Software Engineer at Google working on Gemini for Google Workspace, an agentic LLM assistant across Google Workspace apps such as Docs, Slides, Chat and Gmail.

Prior to that, I obtained my Ph.D. in Computer Science at Arizona State University, where I worked at the Interactive Robotics Lab under the supervision of Prof. Heni Ben Amor. My research focused on Language-Conditioned Robot Learning and LLM Agents, and my publications are shown at NeurIPS, CoRL, COLM, IROS, ICRA, etc. In my PhD research, I mainly focused on 2 topics:

(1) Study of data-efficient imitation learning methods for language-conditioned robot manipulation, by either improving transformer models, or harnessing foundation models for higher data quality.

(2) Leveraging (V)LLMs as a guidance for embodied robot behavior, with the insight that (V)LLMs are capable of providing critiques for robots, maneuvering robot actions for better styles, or even directly optimizing robot policies through in-context linguistic and numerical reasoning.

[ Publications ] [ CV ]

Work Experience

Google LLC

Software Engineer

Aug 2025 - Present

  • Working on Gemini for Google Workspace, a cutting-edge agentic AI assistant which empowers users across multiple Google Workspace apps, including Docs, Slides, Chat, Gmail, etc.
  • Highly involved in data, eval, and quality of this multi-agent LLM harness, i.e., curating evalsets that can reflect the losses from user feedback, implementing eval infra accordingly, and hillclimbing the gaps.
  • Worked closely with first-party teams to onboard new features to this agentic system.

Google LLC

Software Engineering Intern

Summer 2024

  • Proposed and implemented a holistic data quality assurance mechanism for the SFT data of a series of LLMs, for the product Gemini for Google Workspace.
  • The proposed mechanism includes 3 parts: rule-based checkers, LLM-based checkers and AutoEval-based checkers, which identify data format issues (e.g., missing critical fields), semantic issues (e.g., the mismatch between model final response and user query), and overall quality issues (e.g., the overall impression of the final response) respectively.
  • The implemented checkers detected problems up to 20.1% of the training datapoints. This helps the data curation team to improve the data quality, and in turn helps the training of Gemini for Google Workspace LLMs.

LinkedIn Corporation

Machine Learning – Artificial Intelligence Engineering Intern

Summer 2020

  • Created a BERT based LinkedIn member profile modeling approach, namely EntityBERT.
  • Proposed 2 objectives for the training of EntityBERT: masked-entity-modeling and profile-relation-prediction.
  • The trained model was able to predict missing fields of given LinkedIn user profiles, where the top 1 accuracy was increased by over 10% compared to previous baselines.
  • EntityBERT provides the team with a user profile modeling methodology and creates an entity recommendation approach for downstream teams.

Education

Arizona State University
Ph.D. in Computer Science
2021 - 2025

Carnegie Mellon University
M.S. in Artificial Intelligence and Innovation
2019 - 2021

Southwest Jiaotong University
B.E. in Computer Science and Technology
2015 - 2019