Title: Strategic Insufficiency of Consensus: Utilizing Reasoning-Trace Disagreement as a Signal for Knowledge Representation

Abstract:

Standard multi-agent architectures typically prioritize the minimization of discord through mechanisms such as voting, consensus-building protocols, debate structures, or fault-tolerant aggregation. However, we contend that this drive for uniformity is inadequate for tasks involving value judgments, where divergence may signify authentic normative ambiguity rather than mere agent malfunction. Extending previous research on reasoning-trace discrepancies within human-AI collaborative moderation contexts, we introduce a specialized knowledge-representation layer. This layer translates agent decisions and explicit reasoning traces into symbolic disagreement states. Specifically, when agents generate clear reasoning paths alongside binary outcomes, we categorize their interactions into four distinct states based on the intersection of reasoning similarity and conclusion alignment: convergent agreement, divergent agreement, convergent disagreement, and divergent disagreement. These categories facilitate the application of defeasible strategic routing rules. By applying this framework to content moderation, we demonstrate that routing mechanisms sensitive to disagreement can effectively bridge the gap between the sub-symbolic deliberation processes of Large Language Models (LLMs) and symbolic knowledge representation within multi-agent strategic reasoning.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC