Researchers use LLM agents to model viral bank runs in new physics paper

Researchers Chris Ruano and Shreshth Rajan developed an agent-based model where depositors are simulated by a constrained large language model (LLM). The model, validated against lab data, simulates 4,900 bank-run scenarios on a network calibrated to real Twitter activity. It found that within-bank connectivity raises cascade likelihood, and a 10% cross-bank spillover rate triggers a sharp phase transition. The model successfully reproduced the failure order of banks like SVB and First Republic.