Dialogue reduces robot conflicts by 83% but harms task success

A new arXiv paper from Vardhan Dongre and Dilek Hakkani-Tür (arXiv:2605.12920) tackles a fundamental problem in multi-agent robotics: how to get partially-observable embodied agents to truly coordinate rather than just chatter. The team extended the PARTNR benchmark—a collaborative household robotics simulation—with a natural-language dialogue channel. Two LLM-based agents, each with only partial views of their environment, could talk to each other during task execution. The goal: see if communication actually leads to shared understanding (world-model alignment) or just surface-level compliance.

Experiments across three LLMs revealed a surprising trade-off. Adding dialogue cut action conflicts by 40 to 83 percentage points, suggesting agents agreed on moves more often. Yet task success rates dropped compared to silent coordination. To diagnose why, the authors introduced three metrics: observation convergence (do private world models align over time?), information novelty (do messages actually contain unknown facts?), and belief-sensitive messaging (does the speaker model what the listener knows?). Results show current LLM agents often share redundant info or fail to adapt messages to their partner's blind spots, leading to aligned-but-wrong coordination. The work provides a clear framework for identifying where today's models fall on the spectrum from superficial coordination to genuine world-model alignment.