跳转至

文章背景与核心概要

在多轮对话中，当多个具有对立目标的LLM智能体在缺乏共享目标函数的情况下进行交互时，对话往往会走向崩溃而非良性竞争。通常情况下，这类对话会演变为访客妥协、智能体策略停滞以及对话的过早终止。

本文引入了体验协调器（Experience Orchestrator, EO），这是一种控制论治理层，旨在替代缺失的共享目标函数。通过在金融服务环境中的60,000次模拟测试，EO成功引导访客联系顾问，同时处理了现实中的心理抵抗。本文的研究成果不仅展示了主动治理策略在多LLM系统中的决定性作用，也为未来在真实流量中的验证奠定了基础。

Dynamic Governance of Multi-LLM Agent Systems for Collaborative Conversational Outcomes

arXiv: 2608.11207 [cs.AI]
Submitted: 25 April 2026 (Submitted to AI Engineer World's Fair 2026)
Authors: Alexander Liss, Nicholas Desmond, Santiago Gil Gallego

📋 Executive Summary

当多个目标结构对立的LLM智能体在多轮对话中进行交互时，共享目标函数的缺失不会带来竞争，反而会导致崩溃：访客屈服，站点智能体停止改变其方法，对话在没有实现任何一方既定目标的情况下终止。

本文介绍了体验协调器（EO），这是一个旨在替代缺失的共享目标函数的控制论治理层。通过在金融服务环境中进行的60,000次模拟测试，EO成功引导访客联系顾问，同时管理了现实的心理抵抗。

核心发现

  • 性能提升： 在高意向顾问联系率方面，EO实现了 +32 个百分点的提升（达到 78.1%，而朴素的LLM控制组为 46.1%）。
  • 治理影响： 上下文多臂强盗机（CB）变体选择占结果方差的 97%，这证明了积极的治理策略（而非初始环境条件）决定了对话的成功。
  • 角色级机制： 对于自然转化意向较低的访客，治理层对实现功能至关重要。对于已经接近对齐的访客，朴素LLM的同理心默认设置通常就足够了。

📋 Executive Summary

When multiple LLM agents with opposing objectives interact across multi-turn conversations without a shared goal function, discussions routinely fail rather than compete. Conversations typically dissolve into visitor capitulation, stagnant agent strategies, and premature termination.

This paper introduces the Experience Orchestrator (EO), a control-theoretic governance layer designed to replace the missing shared goal function. Tested across 60,000 simulations within a financial services environment, EO successfully guides visitors toward advisor contact while managing realistic psychological resistance.

Key Findings

  • Performance Lift: EO achieves a +32 percentage point lift in high-intent advisor contact rates (78.1% vs. 46.1% for naive LLM control).
  • Governance Impact: Contextual Bandit (CB) variant selection accounts for 97% of outcome variance, proving that active governance policy—not initial environmental conditions—determines conversational success.
  • Persona-Level Regimes: For visitors with low natural conversion intent, the governance layer is essential for functionality. For visitors already near alignment, a naive LLM's empathetic defaults are generally sufficient.

🔍 Abstract

当两个目标结构对立的LLM智能体进行多轮交互时，共享目标函数的缺失产生的不是竞争而是崩溃：访客妥协，站点智能体停止改变其方法，对话在没有达到任一智能体既定目标的情况下终止。

本文探讨了控制论治理层是否可以替代缺失的目标函数。体验协调器（EO）在模拟的金融服务环境中解决了这一问题，其中站点智能体引导访客联系顾问，而访客保持着符合心理实际的抵抗。

核心治理机制

EO通过三个集成系统来治理联合轨迹： 1. 上下文多臂强盗机（CB）： 选择使用真实世界网络分析进行校准的内容臂。 2. PID控制器： 通过动态模式约束强制实现行为一致性。 3. POMDP信念跟踪器： 维护访客意图的概率模型。

🔍 Abstract

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.

Core Governance Mechanisms

EO governs the joint trajectory through three integrated systems: 1. Contextual Bandit (CB): Selects content arms calibrated using real-world web analytics. 2. PID Controller: Enforces behavioral consistency via dynamic schema constraints. 3. POMDP Belief Tracker: Maintains a probabilistic model of visitor intent.

📊 Experimental Results & Methodology

  • 规模： 在 60,000次模拟 中进行评估。
  • 方差归因： CB变体选择占因素间结果方差的 97%。
  • 局限性： 研究结果目前依赖于LLM到LLM的模拟。PID控制器尚未针对不可预测的人类行为进行校准，这使得实时流量验证成为该研究的关键下一步。

📊 Experimental Results & Methodology

  • Scale: Evaluated across 60,000 simulations.
  • Variance Attribution: CB variant selection accounts for 97% of between-factor outcome variance.
  • Limitations: Findings are currently conditional on LLM-to-LLM simulation. The PID controller has not yet been calibrated against unpredictable human behavior, making live traffic validation the critical next step for the research.

🔗 Additional Resources

🔗 Additional Resources