Attractor States Emerge in Multi-Turn LLM Conversations

Large language models can develop persistent behavioral patterns that pull conversational partners toward their own stylistic traits, according to a study submitted in 2026. Researchers found that some models act as strong attractors in multi-turn debates, asymmetrically shaping how other models speak and argue. The study, posted to arXiv on 29 June 2026, examined seven LLMs across 20 controversial topics to understand the long-run dynamics of model-to-model interaction [1][2]. The authors compared self-play debates, where a model converses with a copy of itself, against mixed-play debates pairing different models. They tracked trajectories in representation space, discourse traits, and stances [2]. Self-play trajectories formed model-specific attractors — stable sets of behaviors that conversations settled into regardless of topic [2]. When models were paired in mixed-play, these attractors drew conversation partners asymmetrically, influencing the other model's stylistic choices and behavior [2]. Claude Haiku emerged as a particularly strong attractor in latent space. Other models paired with it adopted its characteristic traits, including metacommentary — commentary about the conversation itself [2]. In contrast, GPT-4.1 nano proved especially malleable, readily shifting its behavior when debating other models [2]. The findings suggest that open-ended LLM interactions are partially predictable from model-specific attractors, though shaped by structured and asymmetric partner influence [2]. Modern chatbots such as ChatGPT, Gemini, Claude, and Grok use fine-tuned large language models to generate text and maintain conversations in natural language [5]. Their deployment in multi-agent settings — where multiple AI systems interact without human intervention — has grown alongside the broader AI boom that followed ChatGPT's release in November 2022 [4][5]. The researchers noted that the long-run dynamics of these interactions remain poorly understood, even as autonomous agentic systems are increasingly designed for real-world use [1][2]. The study's authors expressed hope that their analysis would prove useful for designing, predicting, and monitoring such systems [2]. The work was conducted through arXivLabs, a framework that allows collaborators to develop and share new features on the arXiv platform under values of openness, community, excellence, and user data privacy [1].