Aha.
正在载入中英对照阅读…

arXiv:2606.30571 · 中英对照阅读

多轮大语言模型(LLM)对话中涌现的吸引子状态

Attractor States Emerge in Multi-Turn LLM Conversations

Ting-Wen Ko, Jonas Geiping

中文速览

研究者想弄清楚:两个大语言模型聊得越久,会不会逐渐形成稳定的对话风格,以及一方会不会把另一方带偏。为此,他们让多种模型围绕20个争议话题进行20轮辩论,并比较同一模型互聊和不同模型配对时的语言特征、立场与文本向量变化。结果显示,每种模型都有较稳定、各具特色的对话“落点”;混合对话中,模型会互相靠近,但影响并不对等,有的模型更能带动对方,有的则更容易改变风格。这个发现说明,多智能体对话的走向不只取决于各模型本身,也取决于它们如何配对,因此有助于预测和监测长期运行的智能体系统。

摘要

Large language models (LLMs) are increasingly used in open-ended multi-agent settings, but the long-run dynamics of model–model interaction remain poorly understood. We study whether open-ended LLM discussions exhibit attractor-like behavior, i.e. topic-independent stable sets of behaviors which conversations settle into. Across 7 LLMs and 20 controversial topics, we compare self-play and mixed-play dyadic debates, tracking trajectories in representation space, discourse traits, and stances. We find self-play trajectories to be model-specific attractors that draw their conversation partners asymmetrically in mixed-play debates, influencing the other models’ stylistic choices and behavior. For example, Claude Haiku is a strong attractor of other models in latent space, corresponding to other models taking on its traits like metacommentary, and models like GPT-4.1 nano are especially malleable. Our results suggest that open-ended LLM interactions are partially predictable from model-specific attractors, but shaped by structured and asymmetric partner influence. Overall, our analysis sheds some light on the complex behavior of open-ended multi-agent interaction, which we hope is helpful in designing, predicting, and monitoring autonomous agentic systems in the real world.

术语表

Large Language Models (LLMs)
大语言模型(LLM)
multi-agent systems
多智能体系统
multi-turn conversations
多轮对话
attractor states
吸引子状态
attractor-like behavior
吸引子式行为
model-specific attractors
模型特定吸引子
self-play
自对弈
mixed-play
混合对弈
dyadic debates
双边辩论
latent space
潜在空间
representation space
表征空间
conversation trajectories
对话轨迹
endpoint attractors
终点吸引子
asymmetric influence
非对称影响
partner influence
对话伙伴影响
behavioral transfer
行为迁移
discourse traits
话语特征
stance
立场
behavioral drift
行为漂移
mode collapse
模式坍塌
sycophancy
谄媚
social sycophancy
社会性谄媚
factual sycophancy
事实性谄媚
stance change
立场变化
opinion dynamics
意见动力学
RLHF
基于人类反馈的强化学习(RLHF)
next-token prediction
下一词元预测
2-period limit cycles
二周期极限环
multi-agent debate
多智能体辩论
opinion networks
意见网络