Personalizing Embodied Multimodal Large Language Model Agents over Long-term User Interactions

In three linesPOLAR is a framework for MLLM-based embodied agents that personalizes assistance through a multimodal knowledge graph. It organizes past interactions into semantic memory (visual concepts) and episodic memory (agent trajectories), improving performance especially for multi-hop reasoning and tracking user-specific context updates.

Summary generated by Claude — human-verified