LLM Drift: A Study on Behavioral Drift Phenomenon of Large Language Models in Adversarial Multi-Agent Interactions

Introduction / Main Post: LLM Drift: A Study on Behavioral Drift Phenomenon of Large Language Models in Adversarial Multi-Agent Interactions

This article introduces a research platform for quantifying the LLM Drift phenomenon. The platform uses LangGraph to build adversarial debate simulations and combines 22-dimensional behavioral metrics to evaluate the drift of models across five dimensions—psychometrics, personality traits, emotional states, cognitive structure, and social relationships—during long-term interactions.