LEMON: Zero-Shot Misbehavior Monitoring via Equivariant Attention Map Learning

LEMON implements equivariant learning over attention maps to enable zero-shot misbehavior detection across different LLMs. By treating cross-model attention mismatches as algebraic symmetries, the method avoids the need for expensive re-annotation when deploying monitors on new model architectures.