跳转至

审计多智能体大模型推理树：性能超越多数投票与LLM作为裁判

文章背景与核心概要

随着大语言模型（LLM）的广泛应用，多智能体系统（MAS）被广泛用于增强其推理能力。然而，传统系统往往依赖简单的“多数投票”（Majority Voting）来聚合结果，这种方法不仅丢弃了宝贵的推理结构，还容易陷入“幻觉共识”（Confabulation Consensus）的陷阱——即多个智能体表现出相关的偏误，并一致认同错误的推理依据。

为了解决这一痛点，本文引入了一种名为 AgentAuditor 的全新框架，它突破了基于频率的聚合方式。AgentAuditor 通过构建显式映射智能体推理轨迹中一致与分歧的“推理树”，并在关键分歧点通过“分支级证据对比”来解决冲突，将全局仲裁转化为高效的局部验证。此外，本文还提出了“反共识偏好优化”（ACPO）方法，利用经证据验证的偏好监督来训练仲裁器，从而抵御误导性的多数派线索。在四个 MAS 框架和多个推理基准上的实验表明，AgentAuditor 持续超越了标准的多数投票和 LLM 作为裁判的基线，实现了高达 5% 的绝对准确率提升，同时保持了极高的 Token 计算效率。

arXiv ID: arXiv:2602.09341 [cs.AI]
Authors: Wei Yang, Shixuan Li, Heng Ping, Peiyu Zhang, Paul Bogdan, Jesse Thomason
Submitted: 10 Feb 2026 (Last revised: 3 Sep 2026)
License: CC BY 4.0 license icon

arXiv ID: arXiv:2602.09341 [cs.AI]
Authors: Wei Yang, Shixuan Li, Heng Ping, Peiyu Zhang, Paul Bogdan, Jesse Thomason
Submitted: 10 Feb 2026 (Last revised: 3 Sep 2026)
License: CC BY 4.0 license icon

📌 Summary

Multi-Agent Systems (MAS) are widely used to boost the reasoning capabilities of Large Language Models (LLMs). However, traditional systems often rely on naive majority voting to aggregate results—an approach that discards valuable reasoning structures and remains vulnerable to confabulation consensus (where agents share correlated biases and agree on an incorrect rationale).

This paper introduces AgentAuditor, a novel framework that transcends frequency-based aggregation by: 1. Constructing a Reasoning Tree that explicitly maps out agreements and divergences among agent reasoning traces. 2. Resolving conflicts via branch-level evidence comparison at critical divergence points, turning global adjudication into efficient, localized verification. 3. Proposing Anti-Consensus Preference Optimization (ACPO), which trains the adjudicator using evidence-verified preference supervision to resist misleading majority cues.

Across four MAS frameworks and multiple reasoning benchmarks, AgentAuditor consistently outperforms standard majority voting and LLM-as-Judge baselines, delivering absolute accuracy gains of up to 5% while maintaining token efficiency.

多智能体系统（MAS）被广泛用于增强大语言模型（LLM）的推理能力。然而，传统系统往往依赖简单的多数投票来聚合结果——这种方法丢弃了有价值的推理结构，并且容易受到幻觉共识的影响（即智能体共享相关偏置并就错误的原理达成一致）。

本文引入了 AgentAuditor 这一新颖框架，通过以下方式超越了基于频率的聚合： 1. 构建推理树，显式映射智能体推理轨迹之间的共识与分歧。 2. 在关键分歧点通过分支级证据对比来解决冲突，将全局仲裁转化为高效、局部的验证。 3. 提出反共识偏好优化（ACPO），使用经证据验证的偏好监督来训练仲裁器，以抵御误导性的多数派线索。

在四个 MAS 框架和多个推理基准上，AgentAuditor 持续优于标准多数投票和 LLM 作为裁判的基线，在保持 Token 计算效率的同时，实现了高达 5% 的绝对准确率提升。

🧭 Abstract

Multi-agent systems (MAS) can substantially extend the reasoning capacity of large language models (LLMs). Most MAS frameworks aggregate agent outputs via simple majority voting, discarding the evidential structure of reasoning traces. Majority voting is brittle under confabulation consensus, where agents share correlated biases and converge on the same incorrect rationale. We introduce AgentAuditor, which moves beyond frequency-based aggregation by organizing agent traces into a Reasoning Tree that explicitly represents agreements and divergences in their reasoning. AgentAuditor resolves conflicts by comparing branch-level evidence at critical divergence points, turning global adjudication into efficient, localized verification. We further propose Anti-Consensus Preference Optimization (ACPO), which trains the adjudicator with evidence-verified preference supervision to reduce conformity to misleading majority cues. Across four MAS frameworks and multiple reasoning benchmarks, AgentAuditor consistently improves aggregation performance over majority voting, with gains of up to 5% absolute accuracy while remaining token-efficient.

多智能体系统（MAS）可以大幅扩展大语言模型（LLM）的推理能力。大多数 MAS 框架通过简单的多数投票聚合智能体输出，从而丢弃了推理轨迹的证据结构。在幻觉共识下，多数投票显得非常脆弱，此时智能体会共享相关的偏置并收敛到相同的错误原理。我们引入了 AgentAuditor，它将智能体轨迹组织成一个推理树，显式表示其推理中的共识与分歧，从而超越了基于频率的聚合。AgentAuditor 通过在关键分歧点比较分支级证据来解决冲突，将全局仲裁转化为高效的局部验证。我们进一步提出了反共识偏好优化（ACPO），使用经证据验证的偏好监督来训练仲裁器，以减少对误导性多数派线索的从众性。在四个 MAS 框架和多个推理基准上，AgentAuditor 持续提升了优于多数投票的聚合性能，在保持 Token 计算效率的同时，实现了高达 5% 的绝对准确率提升。