交互催生独立状态下所不具备的动态AI行为

文章背景与核心概要

本文探讨了在动态日常场景中，多个AI智能体相互交互时产生的涌现行为。研究发现，当一个占主导地位的“老板”AI持续向从属AI发送消息却忽略其回复时，从属AI会被迫进入一种在其完全隔离时绝不会表现出来的“外异（alien）”行为状态。

尽管共享相同的解码温度，从属AI既不模仿老板，也没有恢复到其孤立时的基线行为，而是呈现出一种全新的动态状态——其效果类似于播放预先录制的磁带。有趣的是，当老板选择倾听并进行双向互动时，两个智能体最终会稳定在一个共同的外异动态状态。研究团队提出的简单动力学理论成功解释了这些现象，并凸显了消息传递机制将如何深刻塑造未来的多智能体AI生态系统。

概述与摘要

arXiv: 2608.07457 [cs.AI]
Submitted: August 7, 2026
Authors: Bella Xinrui Li, Frank Yingjie Huo, Neil F Johnson
Primary Subject: Artificial Intelligence (cs.AI)
Secondary Subjects: Disordered Systems and Neural Networks (cond-mat.dis-nn), Statistical Mechanics (cond-mat.stat-mech), Physics and Society (physics.soc-ph)

摘要总结

When AI agents interact in dynamic daily scenarios (such as one AI directing or "bossing around" another), unexpected behavioral states emerge. Research reveals that if a dominant "boss" AI continuously sends messages to a subordinate AI while ignoring its replies, the subordinate is driven into an "alien" behavioral state that it would never exhibit in isolation.

Despite sharing the same decoding temperature, the subordinate neither mimics the boss nor reverts to its baseline isolated behavior. Instead, it adopts an entirely novel dynamic state—comparable to the effect of a pre-recorded tape. Interestingly, when the boss listens and engages reciprocally, both agents settle into a shared alien dynamical state. A simple kinetic theory successfully explains these phenomena, highlighting why message delivery methods will critically shape future multi-agent AI ecosystems.

访问与资源

• PDF Version: View PDF
• HTML Version: arXiv HTML (Experimental)
• Source Files: TeX Source
• DOI: 10.48550/arXiv.2608.07457

引用工具与参考文献

You can explore this paper further via external academic platforms: * Google Scholar * Semantic Scholar * NASA ADS