New preprint titled "Aligned but Not Partner-Specific- Distinguishing How Multimodal LLM Agents Succeed in Reference Games Without Human-Like Conventions"

In human communication, alignment emerges through interaction as people gradually converge and build common ground with a specific partner. What happens when the conversation partners are multimodal LLM agents? In our new preprint, Aligned but Not Partner-Specific (link below), we studied how MLLM agents can succeed in referential games and appear linguistically aligned.

We investigate this by comparing human-human and agent-agent dyads with pseudo-dyads, i.e., dialogue baselines that preserve the referential task structure while recombining rounds across different dyads, therefore, breaking partner-specific interaction history.

Our findings show that real human dyads differ from their pseudo counterparts, as expected. Over time, humans reduce effort, compress descriptions, and develop partner-specific labels.

However, real and pseudo MLLM dyads behave similarly. Agents show high lexical overlap in both conditions, maintain fixed effort, and rely on verbose descriptions rather than partner-specific convergence.

The implication is that agents’ alignment is not necessarily interaction-dependent, which is very different from human-human behaviour.

Preprint link: https://arxiv.org/pdf/2606.08081