Research Papers 论文研究 5h ago Updated 1h ago 更新于 1小时前 49

Belief Cascades Drive Persuasion in LLM Agent Networks 信念级联驱动LLM智能体网络中的说服

Introduces a controlled testbed for studying goal-directed persuasion among LLM agents operating within real-world ego-network topologies Persuasion dynamics are shaped by the interaction of network topology, competition, topic, and model prior across four LLM backbones, five graph structures, and 55 policy statements Direct exposure reliably predicts next-round stance change, while peer relays carry smaller but measurable influence, meaning non-persuader agents can still transmit persuasive for 提出首个针对LLM多智能体网络中说服能力的受控测试平台,基于真实世界ego-network拓扑结构 说服动态取决于网络拓扑、竞争环境、话题类型和模型先验四者之间的交互作用 直接暴露可可靠预测下一轮立场变化,同行中继传递的说服影响较小但可测量 仅分析可见文本会遗漏关键信息:计划策略仅部分实现、行动与内容可能脱节、被说服者很少主动声明立场变化 主张将多智能体说服评估为轨迹级和暴露级过程,需结合信念探测、暴露溯源和行动日志进行综合评估

65
Hot 热度
75
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • Introduces a controlled testbed for studying goal-directed persuasion among LLM agents operating within real-world ego-network topologies
  • Persuasion dynamics are shaped by the interaction of network topology, competition, topic, and model prior across four LLM backbones, five graph structures, and 55 policy statements
  • Direct exposure reliably predicts next-round stance change, while peer relays carry smaller but measurable influence, meaning non-persuader agents can still transmit persuasive force
  • Text-only analysis is insufficient: planned strategies are only partially executed, action choices diverge from message content, and persuadees rarely self-report stance shifts that probes detect
  • The authors advocate evaluating multi-agent persuasion as a trajectory- and exposure-level process using belief probes, exposure provenance, and action logs

Why It Matters

As multi-agent LLM systems become increasingly deployed for debate, research coordination, and information mediation, understanding how persuasion actually propagates through agent networks is critical for building reliable and controllable systems. This work provides one of the first empirical frameworks for measuring agent-to-agent persuasion rather than assuming it, offering practitioners concrete evaluation methodologies. The findings also carry implications for understanding information cascades, misinformation spread, and alignment risks in multi-agent deployments.

Technical Details

  • The testbed grounds LLM agent interactions in real-world ego-network topologies, testing across four LLM backbones, five distinct graph structures, and 55 policy statements to ensure broad generalizability
  • Persuasion is measured through belief probes rather than surface-level text analysis, tracking stance changes at the trajectory and exposure level across rounds of interaction
  • The study distinguishes between direct exposure (targeted persuasion) and peer relay effects (indirect influence through non-assigned agents), quantifying both pathways of influence propagation
  • The authors introduce a multi-dimensional evaluation framework combining belief probes, exposure provenance tracking, and action logs to capture the gap between intended persuasive strategy and actual executed behavior
  • Key finding that persuadees rarely self-report stance shifts detected by probes reveals a significant measurement challenge in multi-agent persuasion evaluation

Industry Insight

  • Multi-agent system designers should implement exposure provenance and belief probing rather than relying on message content alone to assess persuasion outcomes, as visible language significantly underreports actual stance movement
  • The peer relay effect means that even agents not explicitly tasked with persuasion can become vectors for influence, suggesting that network topology design is as important as individual agent capabilities in controlling information flow
  • As LLM agent networks are deployed in high-stakes domains like debate, negotiation, and information curation, the gap between planned strategy and executed behavior identified here should be treated as a reliability risk requiring explicit monitoring and intervention mechanisms

TL;DR

  • 提出首个针对LLM多智能体网络中说服能力的受控测试平台,基于真实世界ego-network拓扑结构
  • 说服动态取决于网络拓扑、竞争环境、话题类型和模型先验四者之间的交互作用
  • 直接暴露可可靠预测下一轮立场变化,同行中继传递的说服影响较小但可测量
  • 仅分析可见文本会遗漏关键信息:计划策略仅部分实现、行动与内容可能脱节、被说服者很少主动声明立场变化
  • 主张将多智能体说服评估为轨迹级和暴露级过程,需结合信念探测、暴露溯源和行动日志进行综合评估

为什么值得看

本文首次系统性地量化了LLM智能体网络中的说服机制,填补了多智能体交互研究中"说服能力"评估的空白。对构建可靠的多智能体协作系统、理解AI信息传播动力学具有重要参考价值。

技术解析

  • 实验设计:跨4个LLM骨干模型、5种网络图结构、55个政策声明的受控测试平台,基于真实世界ego-network拓扑构建智能体网络
  • 核心发现:直接暴露(direct exposure)是预测下一轮立场变化的可靠指标;未被分配说服任务的同行智能体仍能通过中继传递可测量的说服影响
  • 方法创新:提出信念探测(belief probes)、暴露溯源(exposure provenance)和行动日志(action logs)三位一体的评估框架,超越传统文本分析
  • 关键洞察:计划策略与实际执行消息存在偏差,行动选择可能与消息内容脱节,被说服者很少主动报告立场变化,说明可见语言不足以反映真实立场移动

行业启示

  • 多智能体系统设计需重视网络拓扑对信息传播的影响,说服效果不仅取决于内容质量,还受连接结构和竞争环境制约
  • 评估AI智能体交互能力时,应超越表面文本分析,采用轨迹级和暴露级的综合评估方法,避免低估或误判实际影响
  • 随着多智能体系统在辩论、研究协调、用户模拟等场景的普及,建立标准化的说服能力基准测试将成为行业基础设施需求

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Agent Agent Research 科学研究 Evaluation 评测