About

I am currently a second-year PhD student in Purdue ECE, advised by Xiaoqian (Joy) Wang, working on LLM interpretability.

I earned B.A. degrees in Computer Science and Mathematics, with a minor in Philosophy, from Boston College. During my undergraduate studies, I was advised by Professors Emily Prud’hommeaux, Sergio Alvarez, Yuan Yuan, and Maira Marques Samary.

Email: zhan5982@purdue.edu

Interests

Awards & Achievements

• John McCarthy, S.J., Award for Natural Science — Boston College, May 2024. Awarded annually to one student for the best senior thesis in Natural Science.
• Neuhauser Award — Boston College Computer Science, May 2024. Awarded annually to 1–2 Computer Science students receiving exceptional recognition from faculty.
• Scholar of the College — Boston College, May 2024. Awarded to exceptional students who completed substantial, high-quality independent work under faculty supervision.

Research Projects

Rethinking SAE Evaluation: An Atomic Interpretable Unit Framework Reveals Hidden Polysemanticity and Redundancy

accepted NeurIPS 2026 · Evaluations & Datasets Track

Riteng Zhang, Xiaoqian Wang

We evaluate sparse autoencoders as systems of atomic interpretable units, examining non-redundancy, completeness, monosemanticity, dead units, and sparsity. Our framework reveals a dominant-feature failure mode: highly influential features can remain polysemantic or redundant even when conventional evaluations appear favorable.

Paper (arXiv) Code SAE checkpoints

Research visual withheld during peer review.

Research title withheld during peer review

in review ICLR 2027

Research summary and project details are withheld while the work is under review.

Title and details withheld during review.

How values may connect to internal features and concrete choices.

What Models Say vs. How They Act: Mechanistic Interpretability of Value–Action Gaps

ongoing research Mechanistic interpretability · language model behavior

We study how language models translate abstract value preferences into concrete choices. The project connects behavioral experiments with activation-level and causal evidence inside the model.

Open the Value Feature Browser

From individual interpretability methods to a reusable formal toolkit.

A Framework for Formalizing and Automating Interpretability

academic work Boston College senior thesis · 2024

My senior thesis introduced the Formalizing and Automating Interpretability Toolkit (FAI), a structured way to describe, compare, select, and implement interpretability methods.

Read the thesis (Google Drive)

Teaching

At Boston College, I served as a teaching assistant for two upper-level computer science courses. I supported Machine Learning for two semesters.