[CS] Synthia Wang Candidacy Exam/Dec 10, 2025

[via cs]

This is an announcement of Synthia Wang's Candidacy Exam.
===============================================
Candidate: Synthia Wang

Date: Wednesday, December 10, 2025

Time: 12 pm CST

Remote Location: https://urldefense.com/v3/__https://uchicago.zoom.us/j/98177134376?pwd=sWk7xY45bs8L3cxgKrSlbgmGihHfg0.1__;!!BpyFHLRN4TMTrA!90mXEa2zTeiW3G5Xda_UOrLX9UuvvPalHkDcpZX2xDWPuevcYPU1BL6wWBHmr8rutSVWwV9DKjbmWrwZpiQ$

Location: JCL 354

Title: Evaluating LLM Support Quality and Privacy Leakage in Mental-Health Conversations

Abstract: LLMs are increasingly positioned as sources of support in sensitive domains, particularly mental health. Their ability to generate fluent, empathetic, and context-aware responses raises both opportunities for accessible assistance and concerns about trust and privacy. Yet, despite rapid advances in conversational modeling and interest in privacy-preserving AI, we lack a systematic understanding of how these systems behave when users disclose personal struggles or seek guidance.

Our work examines LLM behavior in mental-health–related conversations around three themes: (a) support quality, evaluating how LLM responses compare to those written by licensed therapists; (b) user estimation of implicit inference, studying how users anticipate and attempt to mitigate what LLMs can infer; and (c) information leakage, measuring both explicit and implicit personal information exposure in synthetic conversations. We view these themes as fundamental to characterizing the benefits and risks of LLM deployment in high-stakes, sensitive settings, and to guiding the design of safer, privacy-aware mental-health technologies.


Advisor: Nick Feamster

Committee Members: Nick Feamster, Marshini Chetty, Ananya Joshi, Nina Taft