Research

NaturalPRISM: A Natural Language Premise Retrieval Task for Research-level Intermediate Proof States

First author. Under review, MathNLP 2026.

A benchmark of 1.78 million query/answer pairs that match a partial natural-language proof with the premise invoked at its next step, mined from 235,314 arXiv mathematics papers spanning all 32 subject categories. Existing premise-retrieval benchmarks are either built for formal languages like Lean, or are small enough to limit fine-tuning. We fine-tuned Qwen3-Embedding models on domain-specific subsets and found that specialized 4B retrievers outperform generalist models nearly twice their size, and that routing queries to domain specialists through a lightweight classifier performs better still.

Beyond Theorem Proving: Using LLMs to Generate New Conjectures

Sole author. Poster, Joint Mathematics Meetings, Washington, DC, January 2026.

Earlier work from the same project, on whether fine-tuning an LLM's embedding space on mathematical similarity tasks produces a usable notion of "relevant mathematics." Embedded 100,000 arXiv theorems and showed that after fine-tuning, theorems linked by citation move measurably closer together.

Undergraduate AI Researcher, Colorado College

Retrieval and evaluation systems for research-level mathematics: data pipelines for fine-tuning large language models, retrieval-augmented generation for mathematical text, and AI-based evaluation models for assessing the mathematical value of natural-language outputs.

Colorado College, BA in Computer Science with Distinction, minor in Mathematics. May 2026. Grace Hopper Excellence Prize, awarded annually to one graduating senior for exceptional commitment to the computer science community.