Wenpei Shao

I am a Research Data Engineer at the Center for Healthy Minds, University of Wisconsin–Madison, where I work with Ross Jacobucci on reinforcement learning and digital phenotyping: what a learning agent should attend to, and what a person's phone reveals about their mental state.

I received my B.S. in Computer Science from UW–Madison in 2025 and was a visiting student researcher at Westlake University in 2024.

Email / Scholar / GitHub / LinkedIn

Research

I study which signals a learning system should compute over. Much of machine learning fixes the objective and varies the model; my work holds the model fixed and asks what the objective should see. Three threads:

Reward and exploration design for sequential decision-making. When does a reward model need to observe the state rather than only the action, and what should an exploring agent treat as novel? I develop cheap diagnostics that answer these questions before expensive training.
Digital phenotyping. Reading a person's mental state from what their phone already shows them: vision-language models over passively captured screenshots, and language models over their own messages, evaluated against momentary self-report and clinical events.
Personalized language models as measurement instruments. Fine-tuning a model on one person's writing and placing the resulting agent in standardized conversations, to surface suicide risk more sharply than the raw text does.

I work with Ross Jacobucci and Brooke Ammerman on a cohort of adults with recent suicidal ideation.

Publications

	When Does Knowing the State Help? Diagnosing Process vs. Outcome Reward Design Wenpei Shao, Ross Jacobucci NeurIPS, 2026 paper / OpenReview A five-minute, model-free diagnostic (state-lift) that tells you whether a reward model needs to condition on the state, with a regret bound that turns the statistic into a decision threshold, an admissibility audit for the labels, and a matched Llama-8B training study of what the probe does and does not predict.
	Authorship-Resolved Detection of Suicidal Ideation in Smartphone Screenshots: Which Expressions Track Momentary Self-Report and Clinical Events Wenpei Shao, Brooke Ammerman, Ross Jacobucci IEEE ICDM Workshops (ICDMW), 2026 paper Suicide-related text on a phone screen may be written by the owner, received from someone else, or encountered as news or fiction. We assign authorship from interface geometry, merge repeated captures into episodes, and show that only the owner's own active ideation tracks momentary self-report.
	Predicting Momentary Suicidal Ideation From Smartphone Screenshots Using Vision-Language Models: Prospective Machine Learning Study Ross Jacobucci, Wenpei Shao, Veronika Kobrinsky, Brooke Ammerman JMIR Mental Health, 13:e90581, 2026 paper Vision-language models read passively collected smartphone screenshots to predict momentary suicidal ideation in a prospective study.
	Surfacing Suicidal Risk Through Simulated Social Interaction: Per-Person Language Model Agents as Communicative Stress Tests Wenpei Shao, Brooke Ammerman, Ross Jacobucci medRxiv, 2026 (preprint) paper / medRxiv Language-model agents built per participant are placed in simulated social interactions; their responses under communicative stress surface suicidal risk that static measures miss.

Under review

	Counting Capabilities, Not States: Capability Representations for Exploration in Open-Ended Worlds Wenpei Shao, Ross Jacobucci Under review, 2026
	Personalized Language Models Encode Suicidal Ideation Even When They Do Not Express It Wenpei Shao, Brooke Ammerman, Ross Jacobucci Under review, 2026

Experience

2025–	Research Data Engineer, Center for Healthy Minds, University of Wisconsin–Madison
2024	Visiting Student Researcher, Westlake University
2023	Software Engineer Intern, China Merchants Bank
2022–2025	MERIT IT Student Staff, School of Education, University of Wisconsin–Madison

Education

2021–2025

B.S. in Computer Science, University of Wisconsin–Madison

Center for Healthy Minds, 625 W. Washington Ave., Madison, WI 53703 · wshao33@wisc.edu
Template adapted from Jon Barron's website.