arXiv:2605.26256v1 Announce Type: new Abstract: Multimodal large language model (MLLM)-based embodied agents have shown strong potential for solving complex tasks in physical environments. However, personalized assistance requires more than following generic instruction or recogn