LEMON: equivariant attention-map monitor for zero-shot LLM misbehavior (arXiv:2610.22723v1)
The paper introduces LEMON (Language-model Equivariant MOnitoring Network), an online graph neural network that processes attention maps token-by-token and produces invariant predictions under algebraic symmetries across models to enable zero-shot cross-model monitoring of LLM misbehavior. Evaluated on 16 LLMs and six datasets for hallucination and jailbreak detection, LEMON outperforms baselines in the zero-shot setting, improving up to 14 ROC‑AUC points while incurring negligible runtime cost.