Research
My work builds systems that make LLMs useful for expert-driven scientific tasks — with a focus on reliability, explainability, and tight integration with domain knowledge.
Preference Learning from Physics-Based Feedback for Materials Design · Present work
I develop post-training pipelines that use physics-based simulation outcomes as reward signals to tune LLMs toward reliable inverse alloy design. Rather than relying on human preference labels, the system closes the feedback loop with domain-specific evaluators — ensuring generated alloy compositions are physically plausible and satisfy microstructure constraints. This approach was presented as a Spotlight paper at the NeurIPS 2025 AI4Mat Workshop.
This work is supported by the AI2 Open Model Initiative (OMAI) grant (PI: Noah Smith, Co-PI: Samuel Carton).
Scientific Information Extraction from Materials Literature · Present work
I study how reliably LLMs can extract structured knowledge (e.g., material properties, experimental conditions) from scientific text. Working with two real-world materials datasets, I benchmarked extraction accuracy under distribution shift and noisy annotations, and developed calibration methods that improve reliability without requiring full retraining. Published in Findings of ACL 2024.
Sample-Efficient Sentiment Analysis on Social Media · Concluded · 2025
Developed a sample-efficient NLP approach to analyze Twitter opinion on US aid to Ukraine, demonstrating that high-quality sentiment classifiers can be built with substantially reduced annotation budgets. The work contributes methods applicable to low-resource opinion mining in rapidly evolving policy and social contexts. Published at HICSS 2025.
NLP Pipeline for Patient–Trial Matching · Concluded · May 2024
In collaboration with physicians at the Stephenson Cancer Research Center, I built an automated NLP pipeline that maps patient clinical notes to eligibility criteria for early-phase oncology trials. The system used explainable AI components to flag the specific evidence driving each match decision, giving clinicians a transparent and auditable recruitment tool. Published in PLOS ONE (2024).
Conversational Search & Dialogue Systems · Concluded · May 2023
I studied how users formulate search requests as conversational speech acts, and built classifiers that map dialogue turns to structured search actions. This work produced two publications — ACM/IEEE JCDL 2023 (Best Paper Nominee, top 3) and ACM CHIIR 2021 — and contributed frameworks for grounding task-oriented dialogue in information retrieval.