PolicyShiftGuard: benchmarking and improving policy-adaptive image guardrails

The researchers propose the policy-adaptive image guardrail task and release the benchmark PolicyShiftBench and the guardrail model PolicyShiftGuard. The benchmark contains 2,000 policy discrimination instances over 265 images, with an average of 7.55 policy-conditioned prompts per image, testing whether models follow the current policy rather than relying on image-level safety priors. Experiments show that existing VLMs and dedicated guardrails remain fragile under policy switching; PolicyShiftGuard uses two-stage training with randomized policy SFT (RP-SFT) and boundary-pair policy adaptation (BP-Adapt), using a pairwise comparison loss to distinguish block and allow policies. The authors state that its 7B model achieves SOTA 76.9 average F1 and 72.1 average PSS on PolicyShiftBench, and transfers well to UnSafeBench and SafeEditBench; ablations confirm that matched allow/block boundary pairs are key to stable policy adaptation.