Where Reasoning Diverges: Localized Multi-Agent Debate

2026-08-02

AI Summary

Existing multi-agent debate frameworks often exchange entire reasoning chains, even when disagreements pertain only to a few intermediate assertions, resulting in inefficiency. This work proposes Localized Multi-Agent Debate (LMAD), a protocol that models reasoning trajectories as typed node graphs and employs a conflict localization algorithm to identify the earliest point of divergence, enabling agents to debate only the relevant local subgraph. A state-locking mechanism prevents redundant discussion of already-consensus steps. Evaluated on four multi-hop question-answering benchmarks, LMAD achieves the highest macro-average judgment accuracy across ten backbone models, outperforming the strongest baseline by up to 7.20 percentage points, thereby substantially enhancing both the efficiency and accuracy of collaborative reasoning.

Abstract

Multi-agent debate commonly exchanges complete reasoning traces even when disagreements concern only a few intermediate claims. We introduce Localized Multi-Agent Debate (LMAD), an inference-time protocol that represents agent traces as typed nodes, locates their earliest conflict, and restricts debate to the corresponding local segments. Guarded resolution extends a shared committed state so that later conflicts can be addressed without reopening accepted steps. We evaluate a single fixed LMAD configuration on four multi-hop question-answering benchmarks using ten backbones from four model families. Our method achieves the highest macro-averaged judge accuracy across all ten backbones, outperforming the strongest conventional baseline by up to 7.20 percentage points.

Problem

• multi-agent debate
• reasoning traces
• localized reasoning
• conflict resolution
• multi-hop question answering

Innovation

• Localized Multi-Agent Debate
• reasoning traces
• conflict localization
• guarded resolution
• multi-hop question answering