PolicyShiftGuard introduces benchmark and model for policy-adaptive image guardrails

Researchers introduce PolicyShiftBench, a benchmark with 2,000 instances over 265 images, to evaluate whether models adapt to active safety policies rather than relying on image-level priors. They also propose PolicyShiftGuard, a compact 7B model trained via Randomized Policy SFT and Boundary-Pair Policy Adaptation to handle policy shifts.

PolicyShiftBench pairs each image with an average of 7.55 policy-conditioned prompts to test adaptation to held-out policy definitions.
PolicyShiftGuard uses a two-stage training recipe combining RP-SFT with BP-Adapt, which separates blocking and passing policies via pairwise comparison loss.
The model achieves state-of-the-art performance on PolicyShiftBench with 76.9 Avg. F1 and 72.1 Avg. PSS scores.
It transfers well to UnSafeBench and SafeEditBench while improving the latency-performance trade-off through a concise output format.

The authors consider this important because existing VLMs and specialized guardrails remain brittle under policy shifts, whereas PolicyShiftGuard substantially improves policy-sensitive performance.

Benchmarks