Current Work

At NYU, working with Mark Ho and Ilia Sucholutsky, I study how and when people trust and defer to AI, and what makes AI worth deferring to in the first place. The work runs along three threads: trust and deferral in teaching, data extrapolation, and visual categorization. More to come soon.

Doctoral Research

My Ph.D. work at Princeton, advised by Yael Niv and Tom Griffiths and in close collaboration with Mark Ho, developed computational models that distinguish mentalizing from heuristic strategies in human teaching. Across behavioral experiments and Bayesian model comparison, we found strong individual differences: some people teach in ways consistent with costly mentalizing while others rely on low-effort heuristics. People also switch strategies adaptively across environments. A lightweight inference-scaffolding intervention increased mentalizing-consistent teaching, and improved outcomes even after the prompts were removed. This work appeared in Nature Human Behaviour (paper).

My dissertation also carried these models over to large language models. Most models teach well, matching human performance graph by graph, and their choices are best explained by Bayes-optimal teaching rather than by the simpler heuristics many people fall back on. They also change strategy very little from trial to trial. But the nudges that worked on people did not carry over: models follow inference- and reward-focused prompts when asked, yet those scaffolds do not reliably improve their later teaching on the graphs built to catch heuristics, and sometimes make it worse. Prompt compliance is not the same as better teaching. That work, with Mark Ho and Tom Griffiths and senior authors Yael Niv and Ilia Sucholutsky, appeared at the ICLR 2026 workshop From Human Cognition to AI Reasoning (arXiv).

Earlier Research

Working with Angela Langdon, in collaboration with Geoffrey Schoenbaum's lab at the National Institute on Drug Abuse, I contributed computational modeling to work showing that dopamine neuron activity in an odor-guided choice task reflects multiple parallel reward predictions, about both when and what reward is expected, rather than a single value stream (Nature Neuroscience, 2023).

Before graduate school, I studied spatial navigation with Arne Ekstrom and Robert Wilson at the University of Arizona. Using VR and treadmill path-integration tasks, we showed that body-based and landmark cues are combined when they roughly agree but compete when they conflict (PLOS Comp Bio, 2022), that homing errors during blind walking are better captured by vector-addition than encoding-error models (PLOS Comp Bio, 2020), and, in a companion piece, questioned the assumed one-to-one mapping between grid-cell codes and navigation behavior (Hippocampus, 2020).

Earlier still, in Clifford Saron's lab at the UC Davis Center for Mind and Brain, I processed EEG data from autistic children and ran independent component analysis for work showing that children with and without disproportionate megalencephaly have distinct loudness-dependent auditory responses, evidence for a physiologically distinct ASD subtype (Autism Research, 2019).