Bi-Directional Mental Model Reconciliation for Human-Robot Interaction with Large Language Models