Guardian-as-an-Advisor: Advancing Next-Generation Guardian Models for Trustworthy LLMs

The research proposes a soft-gating pipeline, Guardian-as-an-Advisor: the Guardian model gives a risk label and a brief explanation, which are prepended to the original query for re-reasoning, while the base model still operates according to the original model specification, reducing over-refusal while maintaining safety. A companion 208k+ multi-domain dataset, GuardSet, is constructed, including robustness and honesty slices; GuardAdvisor is trained with SFT and RL to ensure consistency between labels and explanations. GuardAdvisor has competitive detection accuracy, its reasoning accounts for less than 5% of the base model's compute, and end-to-end overhead is only 2-10% at realistic harmful input rates.