AI for Behavioral Science

Robin Na

Building and evaluating AI agents that learn from scientific evidence and predict human behavior.

I am a PhD candidate in Information Technology at MIT, advised by Abdullah Almaatouq and Michiel Bakker. My work spans research synthesis, behavioral simulation, and evaluation under distribution shift.

I build systems that turn papers and behavioral data into predictions, and study when those predictions hold up in new settings. I also develop multi-agent environments to evaluate how AI scientists gather evidence and make research decisions.

I previously interned with AI research teams at Amazon in 2023 and 2024, working on social networks, LLMs, and knowledge retrieval. Before MIT, I was a research fellow at Carnegie Mellon University. I hold a B.S. in Physics and a B.A. in Knowledge Ecology from Seoul National University.

Explore my research Curriculum vitae
Robin Na

Research

Designing personas for behavioral simulation

Which features of a person matter for simulating their behavior? I build an automated framework that iteratively constructs behavioral profiles for LLM simulations. It selects features by their ability to predict collective human behavior under experimental designs that differ from those used to develop the profiles.

Designing Behavioral Profiles for Language Model Simulation of Out-of-Distribution Human Behavior

Robin Na, Kehang Zhu, Logan Cross, Michiel Bakker

Under review at Information Systems Research · Accepted at UserSim at NeurIPS

Undocumented Evidence and Collective Failure in Multi-Agent AI Scientists

How should an AI scientist decide what to study next? I build multi-agent research environments to evaluate how agents allocate data-collection budgets and combine heterogeneous, noisy, and potentially misleading evidence. The goal is to understand how those research decisions shape predictions and policy recommendations in new settings.

Robin Na, Michiel Bakker

Working Paper

Can scientific literature improve AI predictions?

I build research-synthesis agents that extract and integrate evidence from scientific papers to predict behavioral experiments. By evaluating predictions on new experimental designs, this work examines both the capabilities of AI research agents and the predictive value of the literature they draw on.

Auditing Scientific Literatures with Large Language Models

Robin Na, Duncan J. Watts, Abdullah Almaatouq

Working Paper · Presented at the AI for Science Workshop at ICML

See my CV for the full list of papers.

Connect

I always love to interact with and learn from people with diverse perspectives. If there are any interesting research ideas you would like to discuss, please don't hesitate to email me!

robinna [at] mit.eduLinkedIn Google Scholar