A study investigates whether open-ended large language model discussions exhibit attractor-like behavior by analyzing trajectories across seven models and twenty controversial topics. The research compares self-play and mixed-play dyadic debates to understand how conversations settle into stable sets of behaviors.
- Self-play trajectories form model-specific attractors that asymmetrically influence partners in mixed-play debates.
- Claude Haiku acts as a strong attractor, causing other models to adopt its traits like metacommentary.
- Models such as GPT-4.1 nano are identified as especially malleable within these interactions.
- Open-ended LLM interactions are partially predictable from model-specific attractors but shaped by structured partner influence.
These findings help in designing, predicting, and monitoring autonomous agentic systems by shedding light on the complex behavior of open-ended multi-agent interaction.