Research

I work at the intersection of computational social science and human-AI interaction. My work combines causal inference, large-scale experiments, and human-centered design, with a recurring focus on heterogeneous effects: who benefits from an intervention, who does not, and why.

Platform Interventions & User Agency

How network structures, algorithmic curation, and platform policies shape user behavior—and how to design socio-technical mechanisms that return meaningful control to users.

Wikipedia newcomer mentorship — Causal analysis of 35K+ first conversations showing a mentor's reply raises retention, most for technical questions. → ICWSM 2027
User agency on social media — Analysis of the Prosocial Ranking Challenge ( $N = 9.3 K$ ) on feed ranking trade-offs, paired with a conversational Bluesky tool that lets users personalize their feeds.

Human-AI Interaction

How interactive AI acts as an external cognitive scaffolding for human reasoning, and how to build systems that support deliberation, perspective-taking, and active participation while preserving independent agency.

AI and future thinking — Operationalizing "future thinking" (prospection) in human-AI dialogue and detecting prospective reasoning at scale across 3M+ conversation logs.
Human-AI co-writing & attitude change — The As-If Machine, showing across 2 RCTs ( $N = 1, 901$ ) that co-writing personalized scenarios increases psychological closeness and policy support[cite: 3, 6]. → Nature (under review)
AI scaffolding for participation — Lookover, a tool helping students join peer discussion while keeping the thinking their own. → UIST 2026
Simulating user response — Early work on modeling user cognition and behavior from minimal data, which informs how I now use model-generated agents in experimental designs. → CogSci 2024

Graph Machine Learning

How information propagates across complex relational structures, and how algorithmic foundations can model network topologies while preserving node distinctiveness and heterogeneity.

Graph neural networks — DDSM, reframing message passing to mitigate over-smoothing, benchmarked against 15 baselines. → SIGKDD 2026