LPG: Balancing Efficiency and Policy Reasoning in Latent Policy Guardrails

• URL: http://arxiv.org/abs/2605.17329v1
• Date: Sun, 17 May 2026 08:35:38 GMT
• Title: LPG: Balancing Efficiency and Policy Reasoning in Latent Policy Guardrails
• Abstract summary: 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.
• Score: 41.04710068888387
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Guardrails are a critical safety layer for modern AI systems, but their operating regime is changing. As LLMs are deployed as customized assistants, safety policies are increasingly specified at inference time by users, organizations, or regulatory contexts. This makes safety enforcement fundamentally dynamic: the guardrail should adapt to changing safety policies without retraining. Yet this requirement creates a fundamental tension: faithfully judging complex policy contexts demands reasoning capability, while practical deployment requires low-latency responses. We introduce Latent Policy Guardrail (LPG), a guardrail framework that learnssemantic latent deliberation over dynamic policies. LPG compresses the internal deliberation needed for intent interpretation and policy grounding into continuous states supervised by decision-relevant semantics. At inference time, it generates only a compact verdict anchored to the violated policy clauses, preserving auditability while avoiding the latency of explicit reasoning. Across policy guardrail benchmarks, LPG-4B reaches 84.5% average safety accuracy and 77.9% F1 by compressing deliberation into just 10 latent tokens, outperforming the strongest dynamic baseline while running roughly 11 times faster than Qwen3-4B-Thinking under the single-sample evaluation setup. Code and data are available at https://github.com/SaFo-Lab/Latent_Policy_Guard.

Related papers

• LLM-as-a-Tutor: Policy-Aware Prompt Adaptation for Non-Verifiable RL [64.4276583943816]
  Reinforcement learning for non-verifiable instruction relies on judges with prompt-specific rubrics as reward signals.<n>While recent methods adapt these rubrics to the evolving policy during training, the training prompts themselves remain static.<n>This static approach often results in a critical misalignment between prompt difficulty and policy capability, leaving the judge unable to recover a discriminative reward signal.
  arXiv  Detail & Related papers  (2026-07-05T17:05:13Z)
• SafePyramid: A Hierarchical Benchmark for In-context Policy Guardrailing [25.10453609307193]
  In real-world applications, guardrails are often expected to identify unsafe user-model interactions according to application-specific safety policies.<n>In this work, we study this setting under the paradigm of in-context policy guardrailing, where guardrails predict safety violations based on policy specifications provided in context.<n>We introduce SafePyramid, a safety benchmark comprising 1,000 multi-turn conversations across 10 domains and 3,000 corresponding application-specific policies, which together contain 61,699 distinct natural-language rules.
  arXiv  Detail & Related papers  (2026-06-29T07:27:22Z)
• 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)
• SingGuard: A Policy-Adaptive Multimodal LLM Guardrail with Dynamic Reasoning [0.0]
  We present textbfSingGuard, a policy-adaptive guardrail model family for safety assessment in multimodal conversations.<n>SingGuard treats the active policy as a runtime input.<n>It checks the target content against the active policy rule by rule and predicts both the safety label and the triggered rule.
  arXiv  Detail & Related papers  (2026-06-22T05:37:43Z)
• Test-Time Gradient Guidance of Flow Policies in Reinforcement Learning [50.738952715864116]
  Expressive continuous control policies form the backbone of advances in scaling imitation learning for simulated and real robot control.<n>We propose QGF (Q-Guided Flow), an RL algorithm that performs policy optimization entirely at test time.<n> Empirically, QGF outperforms prior test-time RL methods on single-task and goal-conditioned offline RL benchmarks.
  arXiv  Detail & Related papers  (2026-06-09T16:45:57Z)
• ConsisGuard: Aligning Safety Deliberation with Policy Enforcement in LLM Guardrails [57.25411733152009]
  Reasoning-based guardrails improve safety moderation by generating explicit rationales before issuing final decisions.<n>A model may recognize a harmful intent in its reasoning but still predict a safe label, or issue an unsafe decision without policy-grounded justification.<n>We propose ConsisGuard, a consistency-aware framework for reasoning-based guardrails.
  arXiv  Detail & Related papers  (2026-05-29T09:42:08Z)
• Selective Safety Steering via Value-Filtered Decoding [54.87935112120107]
  Large language models (LLMs) are trained to align with human values, but their generations may still violate safety constraints.<n>Existing decoding-time steering methods often intervene unnecessarily, modifying generations that would have been safe under the base model.<n>We propose a new test-time steering method designed to reduce such unnecessary interventions while improving the safety of unsafe responses.
  arXiv  Detail & Related papers  (2026-05-14T12:13:08Z)
• PolicyBank: Evolving Policy Understanding for LLM Agents [51.86716874651299]
  PolicyBank maintains structured, tool-level policy insights and iteratively refines them.<n>PolicyBank closes up to 82% of the gap toward a human oracle.
  arXiv  Detail & Related papers  (2026-04-16T20:29:30Z)
• Steering Frozen LLMs: Adaptive Social Alignment via Online Prompt Routing [37.86964394689924]
  We introduce the Consensus Clustering LinUCB Bandit (CCLUB), a unified framework for adaptive social alignment via system-prompt routing.<n>CCLUB employs a conservative consensus clustering mechanism, effectively preventing unsafe generalization across semantically proximal but risk-divergent contexts.<n>Our theoretical analysis yields a sublinear regret guarantee, demonstrating near-optimal performance of CCLUB.
  arXiv  Detail & Related papers  (2026-03-04T08:00:16Z)
• Conformal Policy Control [50.46542384484142]
  We show how to use any safe reference policy as a probabilistic regulator for any optimized but untested policy.<n>Unlike conservative optimization methods, we do not assume the user has identified the correct model class.<n>Our experiments on applications ranging from natural language question answering to biomolecular engineering show that safe exploration is possible from the first moment of deployment.
  arXiv  Detail & Related papers  (2026-03-02T18:54:36Z)
• Training-Free Policy Violation Detection via Activation-Space Whitening in LLMs [21.5603664964501]
  We propose a training-free and efficient method that treats policy violation detection as an out-of-distribution detection problem.<n>Inspired by whitening techniques, we apply a linear transformation to decorrelate the model's hidden activations and standardize them to zero mean and unit variance.<n>On a challenging policy benchmark, our approach achieves state-of-the-art results, surpassing both existing guardrails and fine-tuned reasoning models.
  arXiv  Detail & Related papers  (2025-12-03T17:23:39Z)
• Bounded Robustness in Reinforcement Learning via Lexicographic Objectives [54.00072722686121]
  Policy robustness in Reinforcement Learning may not be desirable at any cost. We study how policies can be maximally robust to arbitrary observational noise. We propose a robustness-inducing scheme, applicable to any policy algorithm, that trades off expected policy utility for robustness.
  arXiv  Detail & Related papers  (2022-09-30T08:53:18Z)
• Minimizing Safety Interference for Safe and Comfortable Automated Driving with Distributional Reinforcement Learning [3.923354711049903]
  We propose a distributional reinforcement learning framework to learn adaptive policies that can tune their level of conservativity at run-time based on the desired comfort and utility. We show that our algorithm learns policies that can still drive reliable when the perception noise is two times higher than the training configuration for automated merging and crossing at occluded intersections.
  arXiv  Detail & Related papers  (2021-07-15T13:36:55Z)
• Lyapunov Barrier Policy Optimization [15.364174084072872]
  We propose a new method, LBPO, that uses a Lyapunov-based barrier function to restrict the policy update to a safe set for each training iteration. Our method also allows the user to control the conservativeness of the agent with respect to the constraints in the environment.
  arXiv  Detail & Related papers  (2021-03-16T17:58:27Z)

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.