LEMON: Equivariant Learning over Attention Maps for Zero-Shot Misbehavior Monitoring on New LLMs

Explain Like I'm Normal

LEMON introduces a way to monitor LLM misbehavior by analyzing internal attention maps rather than just text outputs. Because it uses equivariant learning, the monitor can transfer to brand-new models without needing expensive re-annotation or retraining. This solves the 'alignment tax' problem where safety tools break every time a new model version is released.