About me
I am a second year Ph.D. student in the Allen School of Computer Science and Engineering at the University of Washington, advised by Yulia Tsvetkov. I do research in Natural Language Processing and Cognitive Modeling, with a growing emphasis on applying AI to improve healthcare and wellness outcomes.
I'm particularly interested in using computational methods to model cognitive processes, including how humans reason, communicate uncertainty, and make decisions in complex domains like healthcare. My long-term goal is to build socially and cognitively aligned NLP systems that support safer, more personalized, and equitable care.
Research interests: AI for Health, Safety & Reliability, Proactive Learning, Social Reasoning, and more!
Before grad school, I received my B.S. and M.S.E. at Johns Hopkins with majors in Computer Science, Cognitive Science (linguistics focus), and Applied Mathematics (statistics focus). I worked as a research assistant at the Center for Language and Speech Processing advised by Philipp Koehn and Kenton Murray.
Please contact me at stelli [at] cs.washington.edu if you are interested in my work!
Click here to view my CV (updated July 25)
Current Projects
-
Proactive Reasoning
I'm thinking about how to identify and proactively seek information using LLMs to improve model safety & reliability with statistical guarantee. How to make LLMs ask good questions? How do we model "intuition" in expert domains like medicine?
-
Socially-Intelligent Personalization
Modeling how different social groups express health concerns and interpret medical advice. Aiming to personalize AI systems for more equitable, culturally-aware health communication.
News
-
2025-02
Check out our new paper "PrefPalette: Personalized Preference Modeling with Latent Attributes" to model human preferences with cognitively grounded attributes for more accurate and explainable preference learning.
-
2025-06
Presenting Spurious Rewards at Cohere Labs [YouTube] [Slides].
-
2025-06
Prompt engineering can elicit similar behaviors in models as RLVR does. We show that "Spurious Prompt" can boost Qwen2.5-Math MATH-500 performance by 20% as well‼️ Check out our new blogpost "Spurious Rewards and Spurious Prompts."
-
2025-05
Doing RLVR on incorrect and even random rewards can boost Qwen2.5-Math MATH-500 performance by 20%🤯 We explore how and why this happens in our new paper "Spurious Rewards: Rethinking Training Signals in RLVR" (blogpost).
-
2025-04
Standard privacy evaluations focus mainly on surface-level lexical identifiers, but we show that semantic leakage is a more serious threat🚨. Check out our new paper "A False Sense of Privacy: Evaluating Textual Data Sanitization Beyond Surface-level Privacy Leakage."
-
2025-02
Check out our new paper "ALFA: Aligning LLMs to Ask Good Questions A Case Study in Clinical Reasoning" by decomposing a complex goal into attributes and synthesizing paired data for preference learning.
-
2024-12
Presenting MediQ at Neurips2024 in Vancouver 🍁.
-
2024-11
Presenting ValueScope and Multilingual Abstention at EMNLP2024 in Miami 🌴.
-
2024-09
Joining Meta FAIR as a visiting researcher.