Tag
The paper introduces STEMMA, a multi-agent framework that adversarially probes self-identity consistency in LLMs, motivated by concerns that knowledge distillation may transfer behavioral traits like identity representation from teacher to student models.