Research

We compare prompting, sparse autoencoders, and linear probes for interpreting and steering LLM agents in social simulations. SAEs help inspect internal features, probes offer calibrated control, and stronger prompting is competitive or better on some tasks.

Published open access in Strategic Management Journal, this paper introduces a framework for using LLM-based AI agents as synthetic subjects in strategy research.

We estimate how labor markets in 141 countries are exposed to frontier AI, showing large cross-country variation, gender gaps, adoption correlations, and remittance-linked indirect exposure.