Embodied Multi-Agent Coordination by Aligning World Models Through Dialogue

Key Findings

Three findings, taken together, show that current LLM agents talk without aligning. Conflicts go down, but so does task success, because the channel is dominated by hallucinated references rather than grounded observations.

The conflict–success tradeoff is robust across LLMs

We evaluate two communication architectures on a paired 80-episode subset: SC (synchronous-costed: messaging consumes a step) and ACF (asynchronous cost-free: messaging is free, partner messages auto-inject). Removing the cost of speaking increases dialogue volume by ~3.5× (3.8 → 13.4 turns/episode), but does not recover task success.

The failure is content, not channel

Pooled Δ-align is negative for every dialogue condition (−0.05 to −0.15), suggesting dialogue actively misaligns world models. But this is a measurement artifact of one specific failure mode: hallucinated entity references inflate belief sets and cancel out the alignment that grounded utterances would otherwise produce. When we restrict to grounded mentions only, Δ-align flips positive (+0.06 to +0.19), exactly as the theory predicts.

Removing message cost does not fix this. Under the free architecture (ACF), belief-sensitive messaging actually collapses: status messages that under SC neutrally targeted partner-knowledge gaps become strongly negative (BSM_STATUS = −0.194), meaning agents under-attend to what their partner doesn't know when speaking is cheap.

Three Diagnostic Metrics

Each metric is defined over the agents' private world graphs and probes a different layer of alignment. Together they separate agents that talk from agents that actually build a shared understanding.

Conclusions

We propose a diagnostic framework for measuring world-model alignment in embodied multi-agent dialogue, evaluated across three models, two communication architectures, and multiple prompting policies. In every setting, enabling dialogue reduces task success, and removing the action cost of messaging does not recover performance.

Our metrics localize the failure to dialogue content, not channel: while grounded dialogue would align beliefs as predicted, hallucinated and forecasted references dominate and reverse the gain. Free messaging (ACF) further exposes additional failure modes that cost-constrained settings partially mask: belief-sensitive messaging collapses when speaking is cheap.

These failures reflect a fundamental mismatch between training and deployment. LLMs are trained on text-only, dyadic, single-session dialogue without grounding or cost constraints, whereas embodied multi-agent settings require references to physical entities, partial observability, and persistent shared state. Dialogue policies learned in the former regime transfer poorly to the latter, producing fluent but ungrounded coordination. Off-the-shelf LLMs are simply not trained for this problem space; closing the gap will require coordination policies trained in-domain, not just better prompting of existing chat models.

These results identify dialogue content as the critical bottleneck, while pointing to future design directions our framework can systematically evaluate:

• Persistent common ground. Mechanisms that bind utterances to shared, grounded references rather than free-text generation.
• Partner-belief modeling. Explicit representations of what the partner has and has not observed, conditioning each utterance on that gap.
• Multi-round deliberation. Allowing agents to repair, confirm, and refine references before committing to action.

Citation

@inproceedings{dongre2026embodied,
  title     = {Embodied Multi-Agent Coordination by Aligning
               World Models Through Dialogue},
  author    = {Dongre, Vardhan and Hakkani-Tür, Dilek},
  booktitle = {Proceedings of the 27th Annual Meeting of the Special
               Interest Group on Discourse and Dialogue (SIGDIAL)},
  year      = {2026}
}