基于去身份化理论的多智能体协商干预机制
Deindividuation-based Intervention Mechanism for Multi-Agent Negotiation
杜文聪 1程渤1
作者信息
- 1. 北京邮电大学计算机学院，北京100876
- 折叠
摘要
大型语言模型（LLM）在多智能体协商中存在显著的从众偏差,当智能体处于孤立状态时,有较高概率将正确答案翻转为错误答案,而现有的基于提示工程的干预方法效果有限。本文借鉴心理学中的去身份化理论,提出了一种去身份化干预机制：通过监测协商过程中的观点变化来识别潜在的从众偏差,在检测到观点翻转时触发干预,将多数派和少数派的论点论据匿名化呈现,隐藏支持者数量信息。关键设计在于使用观点改变前一轮的未受从众影响的信息进行重构,从而避免错误信息的级联传播,在减轻群体压力对智能体决策影响的同时保留有价值的信息交换。在Civil Comments内容审核数据集上的实验表明,该方法显著降低了不良翻转率,有效提升了少数派保护率,整体准确率提高了6.5个百分点。本文提出的去身份化干预机制为多智能体协商系统的可靠性提供了新的思路。
Abstract
Large Language Models (LLMs) exhibit significant conformity bias in multi-agent negotiation, with a high probability of flipping correct answers to incorrect ones when agents are in isolated positions, while existing intervention methods based on prompt engineering demonstrate limited effectiveness. Drawing on the deindividuation theory from psychology, this paper proposes a deindividuation intervention mechanism. By monitoring opinion changes during the negotiation process to identify potential conformity bias, the mechanism triggers intervention when opinion flips are detected. The intervention anonymizes and presents arguments from both majority and minority perspectives while hiding supporter count information. The key design is to use uncontaminated information from the round before the opinion change for reconstruction, thereby avoiding the cascade of erroneous information while preserving valuable information exchange and reducing the impact of social pressure on agent decision-making. Experiments on the Civil Comments content moderation dataset show that this method significantly reduces the bad transition rate, effectively improves the minority protection rate, and increases overall accuracy by 6.5 percentage points. The deindividuation intervention mechanism proposed in this paper provides new insights for enhancing the reliability of multi-agent negotiation systems.关键词
多智能体协商/从众偏差/去身份化理论/大型语言模型/内容审核Key words
Multi-agent negotiation/Conformity bias/Deindividuation theory/Large language models/Content moderation引用本文复制引用
杜文聪,程渤.基于去身份化理论的多智能体协商干预机制[EB/OL].(2026-01-09)[2026-04-03].http://www.paper.edu.cn/releasepaper/content/202601-14.学科分类
计算技术、计算机技术
评论