Attractor States Emerge in Multi-Turn LLM Conversations

2026-06-29

AI Summary

This study investigates whether stable, topic-agnostic behavioral patterns—referred to as attractor states—emerge in multi-turn dialogues involving large language models (LLMs) and how these states interact dynamically. Through self-play and cross-play experiments across 20 contentious topics using seven prominent LLMs, the authors combine representational trajectory analysis, discourse feature quantification, and stance tracking to reveal, for the first time, model-specific attractor states in open-ended LLM conversations. The findings demonstrate that certain models, such as Claude Haiku, act as strong attractors capable of significantly steering other models toward adopting their metacommentary style, whereas others, like GPT-4.1 nano, exhibit high plasticity. These results indicate that interactions among LLMs possess a degree of predictability and are governed by asymmetric influence mechanisms.

Abstract

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.

Problem

• attractor states
• multi-turn conversations
• large language models
• multi-agent interaction
• conversation dynamics

Innovation

• attractor states
• multi-agent LLM interaction
• conversation dynamics
• model-specific attractors
• asymmetric influence