RESEARCH · RESEARCH · #2171
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.
KEY POINTS
- 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.
- By making internal‑representation monitors invariant across model symmetries, LEMON can reduce repeated annotation and retraining when new LLMs appear, enabling cheaper online detection of hallucinations and jailbreaks.
WHY IT MATTERS
By making internal‑representation monitors invariant across model symmetries, LEMON can reduce repeated annotation and retraining when new LLMs appear, enabling cheaper online detection of hallucinations and jailbreaks.