Attractor States Emerge in Multi-Turn LLM Conversations

Ting-Wen Ko, Jonas Geiping

Keywords: LLM, multi-agent

TL;DR: Multi-turn LLM conversations converge to stable, model-specific attractors. Mixed-model dialogues usually end between the models’ self-play attractors, revealing asymmetric influence and stronger effects of model identity than topic.

Abstract: Large language models are increasingly used in multi-agent settings, but the long-run dynamics of model--model interaction remain poorly understood. We study whether open-ended LLM discussions exhibit \textit{attractor-like behavior}: stable regions in conversational behavior space. We run controlled dyadic debates among 7 language models on 20 controversial topics, comparing self-play and mixed-play under minimal prompting, and track each conversation with sentence embeddings, topic-centered PCA, LLM-judged discourse traits, and stance annotations. We find that self-play trajectories move toward stable endpoint regions that are specific to each model rather than to the topic, while mixed-play endpoints usually fall between the corresponding self-play regions, revealing asymmetric influence between partners. These results suggest that open-ended LLM discussions converge to predictable conversational regimes governed more by model identity than by topic, with implications for the design and control of multi-model agent systems.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 375