Consensus Energy Minimization: Ensuring Reliable Convergence in Collaborative Deliberation

Yuting Yan, Yinghao Fu, Haozhou Gao, Aoxi Liu, Shuang Li

Keywords: Consensus Energy Minimization, Collaborative Deliberation, Reliable Convergence

Abstract: Multi-round deliberation among heterogeneous agents—whether humans, AI systems, or domain experts—offers opportunities to reduce diagnostic uncertainty through complementary reasoning. Yet such collaboration can also amplify errors if agents prematurely converge on unreliable conclusions. We propose a lightweight monitoring framework, Consensus Energy Minimization (CEM), that regulates collaborative decision-making without requiring domain-specific supervision. CEM formalizes deliberation as a dynamical system, where a confusion-aware consensus energy functional tracks both disagreement and convergence in low-reliability regions. The monitor applies stopping-time rules to either halt, continue, or steer discussion toward an agent’s local expertise, ensuring convergence to high-confidence consensus. We provide theoretical guarantees showing that, under mild reliability assumptions, CEM provably avoids harmful convergence and achieves stability in safe consensus regions. Empirically, we demonstrate the framework on synthetic and real-world classification tasks, where CEM reduces uncertainty and improves joint accuracy across diverse interaction scenarios (ideal, asymmetric, and noisy). Our results highlight that principled monitoring, rather than model accuracy alone, is key to harnessing the benefits of deliberation.

Primary Area: interpretability and explainable AI

Submission Number: 23159