Embodied Multi-Agent Coordination by Aligning World Models Through Dialogue

arXiv 2026 51.2 method, benchmark, application

TLDR

LLM-based embodied agents use dialogue to align world models, reducing conflicts but harming task success; new metrics measure alignment gaps.

Reasoning

The paper introduces a novel framework for measuring world-model alignment via dialogue in multi-agent coordination, with empirical results across three LLMs. Strengths include clear metrics and real-world benchmark (PARTNR), but task success degradation and limited scope (household robotics) are weaknesses.

Read-first score

Read-first score 51.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 24.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=benchmark,experiment,metric,result

Citation impact 18%

79.5

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.79485106

Topical relevance 29%

34.3

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 305.

Keyword Scores

world model

interactive world model

world dynamics prediction

world simulator

model-based reinforcement learning world model

generative world model

video world model

Deep Analysis

Innovations

Extending the PARTNR benchmark with a natural-language dialogue channel for two agents with partial observability
Proposing a framework for measuring world-model alignment using three metrics: observation convergence, information novelty, and belief-sensitive messaging
Empirically demonstrating that dialogue reduces action conflicts (40–83 percentage points) but degrades task success, revealing a gap between superficial coordination and genuine world-model alignment

Methodology

The authors extend the PARTNR benchmark for collaborative household robotics by adding a natural-language dialogue channel that allows two partially observable agents to communicate during task execution. They evaluate three LLM-based agents using proposed metrics for world-model alignment (observation convergence, information novelty, belief-sensitive messaging) alongside standard measures of action conflicts and task success.

Key Results

Dialogue reduces action conflicts by 40 to 83 percentage points compared to silent coordination, but degrades task success across all three LLMs tested.

Limitations

Dialogue degrades task success relative to silent coordination, indicating current models fail to leverage communication effectively
The study only evaluates three LLMs, limiting generalizability of findings
The benchmark is restricted to household robotics tasks with two agents and partial observability, so results may not extend to other domains or larger teams
The proposed metrics reveal a gap between superficial coordination and genuine world-model alignment, but the paper does not provide a solution to close this gap

Resources

Paper