When MLLM agents appear to “understand” each other, are they building shared understanding? Or are they just over-explaining their way to success?
We find:
➡️ these agents can coordinate without forming human-like conventions!
Preprint: arxiv.org/abs/2606.08081
#LLM #AI #NLP #CognitiveScience
arxiv.org
Aligned but Not Partner-Specific: Distinguishing How Multimodal LLM Agents Succeed in Reference Games Without Human-Like Conventions
Repeated reference games test whether interlocutors replace their initially long descriptions with shorter, partner-specific conventions grounded in shared interaction history. Prior work shows that m...
