AI agents' disagreement as a knowledge signal: New paper challenges consensus focus

In multi-agent AI systems, the standard approach is to minimize disagreement through voting, consensus protocols, or debate. But a new paper from Michał Wawer and Jarosław A. Chudziak argues this is strategically insufficient, especially for value-laden tasks like content moderation where disagreement can reflect genuine normative uncertainty rather than agent error. The authors propose a knowledge-representation layer that abstracts reasoning traces and decisions into symbolic disagreement states. They distinguish four states based on reasoning similarity and conclusion agreement: convergent agreement (similar reasoning, same decision), divergent agreement (different reasoning, same decision), convergent disagreement (similar reasoning, different decisions), and divergent disagreement (different reasoning, different decisions). This framework supports defeasible strategic routing rules—for instance, when divergent agreement occurs, the system might escalate to human review because the surface consensus masks underlying value conflicts. The work bridges sub-symbolic LLM deliberation and symbolic knowledge representation, offering a path toward more nuanced multi-agent strategic reasoning.

The paper, accepted to the LAMAS&SR workshop at FLoC 2026 (KR + ICPL + LICS + CP + FSCD), instantiates the framework in content moderation. In that domain, a moderator agent and a policy agent might both flag a post (agreement) but arrive there via different reasoning—one focusing on hate speech, the other on misinformation. Current systems would treat this as a simple consensus, but the paper's approach recognizes the divergent reasoning as a signal to route the case to specialized review. This prevents premature consensus that could ignore important contextual differences. By formalizing disagreement states as symbolic knowledge, the framework enables AI systems to handle normative uncertainty more transparently and strategically, moving beyond brute-force aggregation toward informed deliberation.