ConstitutionMAS-EC: Peer Constitutional Critique for Aligned Emergent Communication in Decentralized Multi-Agent LLMs

We present a novel framework for building multi-agent language-model systems that develop efficient, role-specialized communication while remaining aligned under adversarial task pressures, without central control. Existing methods optimize either multi-agent interaction for reasoning, or principle-based critique via self-critique or single-system pipelines, leaving distributed teams insufficiently addressed. We propose ConstitutionMAS-EC, where specialized agents (retrieval, reasoning, verification) follow a shared constitution and are accountable via peer critique: each turn, an active agent proposes a message which peers evaluate for violations of five principles (honesty, collaboration, safety, efficiency, competence), triggering bounded revise-and-recheck loops. Critiques distill into “lessons learned” conditioning future behavior, yielding emergent optimization where protocols compress while satisfying alignment requirements. Evaluation on HotpotQA demonstrates favorable multi-objective trade-offs: compared to baselines, our system achieves higher logical consistency (+20% absolute in hard settings), lower constraint violations (−50%), and reduced cost (−12.6% tokens) at competitive accuracy, suggesting a scalable route to aligned emergent communication.