用于实现协作式对话结果的多大型语言模型智能体系统的动态管控
Dynamic Governance of Multi-LLM Agent Systems for Collaborative Conversational Outcomes
浏览论文内容
中文总结 AI 辅助
本文针对多LLM智能体对话崩溃问题，提出体验编排器EO管控层，通过三种机制提升顾问联系率，在6万次模拟中取得显著效果，为多智能体协作提供新方案。
中文摘要 AI 辅助
当两个具有结构对立目标的大型语言模型（LLM）智能体进行多轮交互时，由于缺乏共享目标函数，不会产生竞争，反而会导致对话崩溃：访客智能体屈服，站点智能体停止调整策略，对话终止且未达成任何一方的既定目标。本文探究控制理论管控层是否可替代缺失的目标函数。体验编排器（Experience Orchestrator, EO）在模拟金融服务环境中解决该问题，其中站点智能体引导访客联系顾问，而访客保持符合心理现实的抗拒状态。EO通过三种机制管控联合轨迹：上下文多臂赌博机（Contextual Bandit, CB），其选择的内容臂基于真实网络分析校准；PID控制器，通过动态模式约束强制行为一致性；部分可观察马尔可夫决策过程（POMDP）信念跟踪器，维护访客意图的概率模型。在6万次模拟中，EO使高意图顾问联系率提升了32个百分点（对比朴素LLM对照组，EO的高意图顾问联系率为78.1%，对照组为46.1%），且CB变体选择解释了97%的因素间结果方差，证实管控策略而非环境初始条件决定轨迹最终走向。角色层面分析揭示两种不同场景：对于无自然转化倾向的访客，管控层是系统能否正常运行的关键；对于已接近达成一致的访客，朴素LLM的共情默认设置基本足够。所有发现均基于LLM间模拟，PID控制器未针对真实人类的不可预测性进行校准，在真实流量上验证EO是关键的下一步。
英文摘要
When two LLM agents with structurally opposed objectives interact across multiple turns, the absence of a shared goal function produces not competition but collapse: the visitor capitulates, the site agent stops varying its approach, and the conversation terminates without achieving either agent's stated objective. This paper asks whether a control-theoretic governance layer can substitute for that missing goal function. The Experience Orchestrator (EO) addresses this in a simulated financial services environment where a site agent guides a visitor toward advisor contact while the visitor maintains psychologically realistic resistance. EO governs the joint trajectory through three mechanisms: a Contextual Bandit (CB) that selects content arms calibrated from real-world web analytics, a PID controller that enforces behavioral consistency via dynamic schema constraints, and a POMDP belief tracker that maintains a probabilistic model of visitor intent. Across 60,000 simulations, EO achieves a +32 percentage point lift in high-intent advisor contact rate (78.1% vs. 46.1% over a naive LLM control), with CB variant selection accounting for 97% of between-factor outcome variance -- confirming that the governance policy, not environmental initial conditions, determines where trajectories end up. Persona-level analysis reveals two distinct regimes: for visitors with no natural inclination toward conversion, the governance layer is the difference between a functional system and a non-functional one; for visitors already near alignment, a naive LLM's empathetic defaults are largely sufficient. All findings are conditional on LLM-to-LLM simulation. The PID controller has not been calibrated against real human unpredictability, and validating EO on live traffic is the critical next step.
