I study how language models represent what they can and cannot answer, and how those internal signals shape their behavior. My work combines representation analysis and causal interventions to understand LLM reliability. I also work on AI auditing and computational linguistics.
LLMs encode when math and code questions have no valid answer, yet this signal is poorly aligned with safety refusal. Steering the recognition direction changes whether models acknowledge the invalidity.
Distance from answerable reference representations provides a pre-generation signal of mathematical unanswerability without labeled failure examples. The signal is weaker in code and does not reliably separate factual prompts.
2027 (expected)
M.S. in Computer Science
2023
B.A. in Linguistics & Computer Science
Mar 2026 – Present
FORTIS Lab · University of Southern California
Research on answerability and refusal in LLMs, mentored by Xiyang Hu (Arizona State University). Since June 2026, also working on auditing AI execution traces, supervised by Yue Zhao (USC).
Jan 2026 – Jul 2026
LLM reliability via representation geometry
Developed an unsupervised, single-pass signal of answerability from hidden-state geometry, with matched experiments across math, factual, and code prompts.
Dec 2023 – Dec 2025
Huachuan Die Casting · Chengdu, China
Led the architecture and delivery of a factory-wide manufacturing execution system, working with two engineers to support 100+ users across 10 departments.
May 2023 – Nov 2023
ByteCommerce LLC · Los Angeles, CA
Led educational software projects and taught Python to student clients.
Aug 2020 – Oct 2020
Antigravity Investments Firm · Berkeley, CA (remote)
Applied statistical models to health-tracking data to study correlations.