Research

Research programs spanning safe NLP, cognition, and mental-health support.

Research

My work clusters around three themes-Core NLP & Model Ops, AI ↔ Cognition, and AI for Mental Health-and these pages go deeper into each. The research page below summarizes the most current projects, the recent completions, pre-PhD work, and associated patents.

Ongoing projects

  • Naïve Scientific Misconceptions in Large Language Models - Investigating if and how LLMs internalize childish or naive theories, especially when primed with misleading data.
  • Understanding Graphical Perception in Data Visualization - Extending the arXiv preprint (2411.00257) into a formal evaluation suite that compares human visual cognition to vision-language models and develops better prompts for graphs.
  • AI Patient Bank - Building state-based simulated patients grounded in therapeutic source texts and vetted via evidence-based testing (collaborators: Nathan Paek, William Fang, Hercy Shen, Declan Grabb, Emma Brunskill, Diyi Yang, Ryan Louie).
  • Ask Before You Summarize: Clarification-Driven Summaries from Dialogue Transcripts - Admission of uncertainty and clarification elicitations for richer automated summarization (with Han-Chin Shing, Lei Xu, Joseph Paul Cohen, Jack Moriarty, Chaitanya Shivade).
  • Responsible Evaluation of AI for Mental Health - Drafting standardized evaluation protocols before deployment; see the preprint 2602.00065.

Recent completions (since joining the PhD)

  • On the use of PLMs for cognitive science (preprint 2501.12651) - Framing large PLMs as computational models of human cognition.
  • The Unlearning Mirage: A Dynamic Framework for Evaluating LLM Unlearning (COLM submission, paper exW2SFJK4H) - Stress-testing unlearning with multi-hop probes and activation analysis.
  • A Neural Network Model of Complementary Learning Systems (preprint 2507.11393) - Pattern separation/completion for continual learning.
  • The World According to LLMs (preprint 2508.05525) - How geographic origin biases entity deduction.
  • Can LLM-Simulated Practice and Feedback Upskill Human Counselors? (preprint 2505.02428) - A randomized study with 90+ novice counselors.
  • Do LLMs Suppress Naïve Theories? (OpenReview paper mMSxCWDkzP) - Scientific reasoning in GPT-4o.
  • TN-Eval: Rubric and Evaluation Protocols for Behavioral Therapy Notes (preprint 2503.20648).
  • Development of Cognitive Intelligence in Pre-trained Language Models (EMNLP 2024; preprint 2407.01047).
  • BabyLM Turns 3: Call for papers (preprint 2502.10645).
  • LLMs Assist NLP Researchers: Critique Paper (Meta-)Reviewing (EMNLP 2024 meta paper; preprint 2406.16253).
  • From Intentions to Techniques: A Comprehensive Taxonomy and Challenges in Text Watermarking (Findings of ACL 2025; preprint 2406.11106).
  • How Well Do Deep Learning Models Capture Human Concepts? The Case of the Typicality Effect (CogSci 2024; preprint 2405.16128).
  • Incremental Comprehension of Garden-Path Sentences by LLMs (CogSci 2024; preprint 2405.16042).
  • Multi-Level Feedback Generation with LLMs for Empowering Novice Peer Counselors (ACL 2024; preprint 2403.15482).
  • Natural Mitigation of Catastrophic Interference (ECAI 2024; preprint 2401.10393).
  • What Makes Digital Support Effective? How Therapeutic Skills Affect Clinical Well-Being (Preprint 2312.10775).
  • Pre-training LLMs using human-like development data corpus (BabyLM 2023; preprint 2311.04666).
  • Numeric Magnitude Comparison Effects in LLMs (Findings of ACL 2023; preprint 2305.10782).
  • Helping the Helper: Supporting Peer Counselors via AI-Empowered Practice and Feedback (Preprint 2305.08982).
  • Modeling Motivational Interviewing Strategies on a Peer-to-Peer Platform (CSCW 2022; preprint 2211.05182).
  • When Flue Meets Flang: Benchmarks and LLMs for the Financial Domain (EMNLP 2022; preprint 2211.00083).
  • JARVix at SemEval-2022 Task 2: It Takes One to Know One? (SemEval 2022 paper [ISBN?], see ACL Anthology https://aclanthology.org/2022.semeval-1.19/).

Pre-PhD work

  • CTI-Twitter: Gathering cyber threat intelligence from Twitter using integrated supervised and unsupervised learning (IEEE BigData 2020).
  • Bitcoin Data Analytics: Scalable techniques for transaction clustering and embedding generation (COMSNETS 2021).

Patents

  • Guiding a user to interact with an intelligent computing system using best practices (US20250196874A1).

Want to explore a collaboration? Reach out on Discord or email-happy to connect over any of the themes above.