Research

Artificial Intelligence

Large Language Models (LLMs) have revolutionized the way we solve problems in NLP. My research concerns how we can apply them, responsibly and with measurement rigor, to education and assessment: from scoring what students write, to giving them feedback, to estimating how hard a test question is before a single student has seen it.

Automated Scoring Systems: As LLMs are pretrained on language, they can be effectively fine-tuned to produce scores more accurately than previous methods. I study this across essays and short constructed responses in subjects such as math and science, including ensembles of pretrained models, retrieval-augmented generation for short-answer scoring, and combining human and AI-generated training data. I also care about what it takes to deploy these systems in practice: efficient models that run without large GPU budgets, and confidence modeling that decides which responses a machine can score and which should go to a human in hybrid scoring systems.
Long-Context Scoring: Essays by older students often exceed the input limits of standard transformer models, and the common fix, truncating the text, means the model never sees the full evidence of writing quality. My recent work compares long-context architectures, including Longformer, ModernBERT, and linear-time state-space models such as Mamba, and examines how truncation affects both scoring accuracy and construct validity, as well as inference cost as essays get longer.
Automated Writing Evaluation: How can we use a combination of LLMs and datasets to effectively provide feedback to students as to how to improve their writing? This includes annotating the argumentative elements of an essay and flagging spelling and grammar errors, now with a single generative model fine-tuned with parameter-efficient adapters that reaches human-level accuracy at a fraction of the computational cost of much larger models. These annotations can also feed back into scoring models to improve their accuracy.
Grammatical Error Correction: Given a grammatically incorrect text, how do we rewrite it in a way that is grammatically correct while maintaining the author’s original intent.
Item Difficulty and Simulated Students: Estimating the difficulty and discrimination of a test item has traditionally required expensive field testing. I am developing methods that fine-tune LLMs to act as simulated students at different ability levels, and then reconstruct item characteristic curves from the responses they produce. Related projects align student simulators with item response theory using preference optimization, and extend the approach to multimodal items that combine text with figures, charts, and other images, such as math items and multiple-choice items modeled with 3PL and multiple-choice models.
Response Validity and Student Safety: Not every response is a genuine attempt, and some contain content that signals a student may be at risk. I build classifiers that detect invalid or off-topic responses in automated essay scoring, and models that detect alarming student responses in both text and speech, so that these cases are routed appropriately instead of receiving an ordinary score.
Bias and Validity: How does one prove, statistically, that automated scoring models give higher scores for members of a certain subgroup? How do we also show that the models are giving valid interpretations of the constructs on which the scores are modeled. My work here includes subgroup analyses of scoring differences, probing whether model representations encode demographic information, and linking a model’s hidden states to interpretable features of writing.

More broadly, I work to connect AI and educational measurement as a community, as a co-author of a review of the opportunities and ethical challenges of AI in educational measurement and as a co-editor of the proceedings of the Artificial Intelligence in Measurement and Education (AIME-Con) conference.

Systems Biology

My research concerns systems biology, which is an interdisciplinary field of study that focuses on complex interactions within biological systems. In particular, proteins define a signaling network, where the majority of physical interactions between proteins is tested by co-precipitation.

The big open question, when we interpret the protein-protein interaction network as a graph: what insights do the topological features of the graph tell us about their underlying biological functions? Graph neural networks are a natural way to model protein-protein interaction networks in a way that implicitly accounts for the structure of the network.

Mathematics

Discrete Painlevé equations are discrete versions of the classical Painlevé equations. They are second order non-autonomous recurrence relations with the singularity confinement property, which is a discrete analogue of the Painlevé property: solvability via associated linear problems. Their spaces of solutions are described by rational surfaces that coalesce to elliptic surfaces in a particular limit, hence the classification of discrete Painlevé equations closely follows the classification of elliptic surfaces. That is to say, their symmetries are described by affine Weyl groups, with the master case being associated with a symmetry of affine type \(E_8^{(1)}\).