Zero-Shot Recommendations with Pre-Trained Large Language Models for Multimodal Nudging