Aligned but Not Partner-Specific: How Multimodal LLM Agents Succeed in Reference Games Without Forming Conceptual Pacts
The study tested if multimodal LLM agents coordinate by developing partner-specific reference conventions or just through shared training data. It found that MLLM dyads successfully coordinated without forming history-dependent conventions, maintaining high effort and lexical overlap. This suggests agents rely on verbose description rather than human-like compact referring expressions to achieve task success.