Ting-Wen Ko

Attractor States Emerge in Multi-Turn LLM Conversations

Ting-Wen Ko, Jonas Geiping

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 o…Read more