PolicyShiftGuard: Benchmarking and Improving Policy-Adaptive Image Guardrails

• URL: http://arxiv.org/abs/2607.05910v1
• Date: Tue, 07 Jul 2026 07:04:28 GMT
• Title: PolicyShiftGuard: Benchmarking and Improving Policy-Adaptive Image Guardrails
• Abstract summary: We study policy-adaptive image guardrailing, where a model must decide whether an image violates the currently supplied policy.<n>We introduce PolicyShiftBench, a comprehensive benchmark with 2,000 policy-discriminative instances over 265 images.<n>We then propose PolicyShiftGuard, a compact policy-conditioned guardrail trained with a two-stage training recipe.
• Score: 24.489215914160813
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: 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.

Related papers

• PolicyMem: Geometric Policy Memory for LLM Governance [72.04962480384322]
  We introduce PolicyMem, a geometric policy memory that externalizes natural-language policies as reusable geometric memory objects.<n>A memory writer compiles natural-language policies into policy memory slots, and query-response pairs read the policy memory through projection energy.<n>The resulting policy-evidence profile directly mediates the safety verdict and is reused for policy attribution and post-intervention verification.
  arXiv  Detail & Related papers  (2026-09-12T06:07:43Z)
• PolicyAlign: Direct Policy-Based Safety Alignment for Large Language Models [19.807112285121715]
  In real-world deployment, emerging safety requirements are often specified as natural-language policies.<n>This creates a mismatch between rapidly evolving safety policies and conventional data-driven alignment methods.<n>We propose PolicyAlign, a framework for directly aligning large language models with safety policies.
  arXiv  Detail & Related papers  (2026-06-24T06:10:33Z)
• LPG: Balancing Efficiency and Policy Reasoning in Latent Policy Guardrails [41.04710068888387]
  We introduce Latent Policy Guardrail (LPG), a guardrail framework that learnssemantic latent deliberation over dynamic policies.<n>Across policy guardrail benchmarks, LPG-4B reaches 84.5% average safety accuracy and 77.9% F1 by compressing deliberation into just 10 latent tokens.
  arXiv  Detail & Related papers  (2026-05-17T08:35:38Z)
• Approximate Next Policy Sampling: Replacing Conservative Target Policy Updates in Deep RL [2.2592563948275486]
  To improve a policy, the value function must be accurate on the state-visitation distribution of the updated policy.<n>Conservative updates solve this problem, but at the cost of shrinking the policy update.<n>This paper explores an alternative solution, Approximate Next Policy Sampling (ANPS)<n>ANPS is satisfied if the distribution of the training data approximates that of the next policy.
  arXiv  Detail & Related papers  (2026-05-06T22:02:35Z)
• Towards Policy-Adaptive Image Guardrail: Benchmark and Method [21.041111216560545]
  Vision-language models (VLMs) offer a more adaptable and generalizable foundation for dynamic safety guardrails.<n>Existing VLM-based safeguarding methods are typically trained and evaluated under only a fixed safety policy.<n>We introduce SafeGuard-VL, a reinforcement learning-based method with verifiable rewards for robust unsafe-image guardrails.
  arXiv  Detail & Related papers  (2026-03-01T18:59:21Z)
• Towards Policy-Compliant Agents: Learning Efficient Guardrails For Policy Violation Detection [25.53228630260007]
  PolicyGuardBench is a benchmark of about 60k examples for detecting policy violations in agent trajectories.<n>PolicyGuard-4B is a lightweight guardrail model that delivers strong detection accuracy across all tasks.<n>Together, PolicyGuardBench and PolicyGuard-4B provide the first comprehensive framework for studying policy compliance in web agent trajectories.
  arXiv  Detail & Related papers  (2025-10-03T20:03:19Z)
• EXPO: Stable Reinforcement Learning with Expressive Policies [74.30151915786233]
  We propose a sample-efficient online reinforcement learning algorithm to maximize value with two parameterized policies.<n>Our approach yields up to 2-3x improvement in sample efficiency on average over prior methods.
  arXiv  Detail & Related papers  (2025-07-10T17:57:46Z)
• SPoRt -- Safe Policy Ratio: Certified Training and Deployment of Task Policies in Model-Free RL [54.022106606140774]
  We present theoretical results that place a bound on the probability of violating a safety property for a new task-specific policy in a model-free, episodic setting.<n>This bound can be applied to temporally-extended properties (beyond safety) and to robust control problems.<n>We present experimental results demonstrating this trade-off and comparing the theoretical bound to posterior bounds derived from empirical violation rates.
  arXiv  Detail & Related papers  (2025-04-08T19:09:07Z)
• Convergence of Policy Mirror Descent Beyond Compatible Function Approximation [66.4260157478436]
  We develop theoretical PMD general policy classes where we strictly assume a weaker variational dominance and obtain convergence to the best-in-class policy.<n>Our main notion leverages a novel notion induced by the local norm induced by the occupancy- gradient measure.
  arXiv  Detail & Related papers  (2025-02-16T08:05:46Z)
• Last-Iterate Convergent Policy Gradient Primal-Dual Methods for Constrained MDPs [107.28031292946774]
  We study the problem of computing an optimal policy of an infinite-horizon discounted Markov decision process (constrained MDP) We develop two single-time-scale policy-based primal-dual algorithms with non-asymptotic convergence of their policy iterates to an optimal constrained policy. To the best of our knowledge, this work appears to be the first non-asymptotic policy last-iterate convergence result for single-time-scale algorithms in constrained MDPs.
  arXiv  Detail & Related papers  (2023-06-20T17:27:31Z)
• Memory-Constrained Policy Optimization [59.63021433336966]
  We introduce a new constrained optimization method for policy gradient reinforcement learning. We form a second trust region through the construction of another virtual policy that represents a wide range of past policies. We then enforce the new policy to stay closer to the virtual policy, which is beneficial in case the old policy performs badly.
  arXiv  Detail & Related papers  (2022-04-20T08:50:23Z)
• BRPO: Batch Residual Policy Optimization [79.53696635382592]
  In batch reinforcement learning, one often constrains a learned policy to be close to the behavior (data-generating) policy. We propose residual policies, where the allowable deviation of the learned policy is state-action-dependent. We derive a new for RL method, BRPO, which learns both the policy and allowable deviation that jointly maximize a lower bound on policy performance.
  arXiv  Detail & Related papers  (2020-02-08T01:59:33Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.