SPASM: Solving Persona Drift and Role Confusion in Multi-Turn LLM Agent Simulations

arXiv·medium signal

Luo and Laban address a critical failure mode in LLM-LLM conversations: persona drift, role confusion, and 'echoing' where one agent gradually mirrors its partner across long multi-turn horizons. SPASM provides stable persona-driven agent simulation for generating synthetic training and evaluation dialogues. Directly relevant for teams using LLM agents for synthetic data generation, tutoring systems, or customer support simulations where maintaining consistent personas across hundreds of turns is essential.