PolicyShiftGuard: Benchmarking and Improving Policy-Adaptive Image Guardrails

Hugging Face Daily Papers Papers
image-guardrails safety policy-adaptation benchmarking vlm fine-tuning

Summary

This paper introduces PolicyShiftBench, a benchmark for policy-adaptive image guardrails, and PolicyShiftGuard, a compact model trained with a two-stage method that improves performance under shifting safety policies.

Image guardrails are typically trained and evaluated under a fixed safety policy, implicitly treating safety as an intrinsic property of an image. Real deployments are different: the same image may be allowed in one product, restricted in another, and newly disallowed when a policy boundary changes. We study policy-adaptive image guardrailing, where a model must decide whether an image violates the currently supplied policy and generalize to held-out policy definitions. We introduce PolicyShiftBench, a comprehensive benchmark with 2,000 policy-discriminative instances over 265 images, where each image is paired with 7.55 policy-conditioned prompts on average to test whether models adapt to the active policy rather than relying on image-level safety priors. We then propose PolicyShiftGuard, a compact policy-conditioned guardrail trained with a two-stage training recipe that combines Randomized Policy SFT (RP-SFT) with Boundary-Pair Policy Adaptation (BP-Adapt). BP-Adapt trains matched prompts for the same image and risk category using standard label supervision and a pairwise comparison loss that separates blocking policies from passing policies. Experiments show that existing VLMs and specialized guardrails remain brittle under policy shifts, while PolicyShiftGuard substantially improves policy-sensitive performance. The 7B model achieves SOTA performance of 76.9 Avg. F1 and 72.1 Avg. PSS on PolicyShiftBench, transfers well to UnSafeBench and SafeEditBench, and improves the latency-performance trade-off with a concise output format. Ablations confirm that matched pass/block boundary pairs are essential for stable policy adaptation.
Original Article
View Cached Full Text

Cached at: 07/16/26, 05:42 AM

Paper page - PolicyShiftGuard: Benchmarking and Improving Policy-Adaptive Image Guardrails

Source: https://huggingface.co/papers/2607.05910

Abstract

Imageguardrailsaretypicallytrainedandevaluatedunderafixedsafetypolicy,implicitlytreatingsafetyasanintrinsicpropertyofanimage.Realdeploymentsaredifferent:thesameimagemaybeallowedinoneproduct,restrictedinanother,andnewlydisallowedwhenapolicyboundarychanges.Westudypolicy-adaptiveimageguardrailing,whereamodelmustdecidewhetheranimageviolatesthecurrentlysuppliedpolicyandgeneralizetoheld-outpolicydefinitions.WeintroducePolicyShiftBench,acomprehensivebenchmarkwith2,000policy-discriminativeinstancesover265images,whereeachimageispairedwith7.55policy-conditionedpromptsonaveragetotestwhethermodelsadapttotheactivepolicyratherthanrelyingonimage-levelsafetypriors.WethenproposePolicyShiftGuard,acompactpolicy-conditionedguardrailtrainedwithatwo-stagetrainingrecipethatcombinesRandomizedPolicySFT(RP-SFT)withBoundary-PairPolicyAdaptation(BP-Adapt).BP-Adapttrainsmatchedpromptsforthesameimageandriskcategoryusingstandardlabelsupervisionandapairwisecomparisonlossthatseparatesblockingpoliciesfrompassingpolicies.ExperimentsshowthatexistingVLMsandspecializedguardrailsremainbrittleunderpolicyshifts,whilePolicyShiftGuardsubstantiallyimprovespolicy-sensitiveperformance.The7BmodelachievesSOTAperformanceof76.9Avg.F1and72.1Avg.PSSonPolicyShiftBench,transferswelltoUnSafeBenchandSafeEditBench,andimprovesthelatency-performancetrade-offwithaconciseoutputformat.Ablationsconfirmthatmatchedpass/blockboundarypairsareessentialforstablepolicyadaptation.

View arXiv pageView PDFProject pageGitHub20Add to collection

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2607.05910 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2607.05910 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2607.05910 in a Space README.md to link it from this page.

Collections including this paper1

Similar Articles

SafePyramid: A Hierarchical Benchmark for In-context Policy Guardrailing

Hugging Face Daily Papers

SafePyramid is a hierarchical benchmark with 1,000 multi-turn conversations across 10 domains and 3,000 policies to evaluate guardrail systems' ability to identify safety violations via in-context policy specification. Tests on 10 frontier LLMs show that even GPT-5.5 only correctly identifies all violated rules 54% of the time at the easiest level, highlighting the challenge of reliable in-context policy guardrailing.

Benchmarking Open-Source Safety Guard Models: A Comprehensive Evaluation

arXiv cs.CL

This paper presents a comprehensive evaluation of 14 open-source safety guard models on a curated benchmark of 79,331 samples across 8 NIST safety categories, finding that model size does not correlate with detection performance and that Qwen Guard (4B) achieves the highest recall.