SingGuard: A Policy-Adaptive Multimodal LLM Guardrail with Dynamic Reasoning
2026-06-22 • Computer Vision and Pattern Recognition
Computer Vision and Pattern RecognitionComputation and Language
AI summaryⓘ
The authors introduce SingGuard, a tool that helps check if conversations involving images and text follow safety rules, even when those rules change over time. SingGuard can quickly or carefully decide if something is safe by looking at each rule and explaining which one was triggered. They also created SingGuard-Bench, a large test set with many examples showing tricky safety cases in multimodal chats. Their tests show SingGuard works better than other methods, especially when rules change during use. This makes SingGuard useful for keeping AI conversations safe in different real-world situations.
vision-language modelsmultimodal safetypolicy adaptationreinforcement learningmultimodal question answeringguardrailscross-modal compositionsafety assessmentbenchmarknatural language rules
Authors
SingGuard Team
Abstract
Vision-language models (VLMs) are increasingly deployed in consumer, medical, financial, and enterprise applications. This broad deployment expands the safety surface: risks can arise from multimodal question answering, assistant responses, and cross-modal composition, while moderation policies may vary across products, regions, and deployment stages. Most existing guardrails either rely on fixed taxonomies or target only a narrow set of interaction settings, which limits their adaptability when safety rules change at deployment time. We present \textbf{SingGuard}, a policy-adaptive multimodal guardrail model family for safety assessment in multimodal conversations. SingGuard treats the active policy as a runtime input: given natural-language rules, it checks the target content against the active policy rule by rule and predicts both the safety label and the triggered rule. To balance efficiency and interpretability, SingGuard supports fast, hybrid, and slow inference regimes along a fast-to-slow reasoning spectrum, ranging from direct safety judgments to policy-grounded deliberation. We further optimize this behavior with fast--slow decoupled reinforcement learning. We also introduce \textbf{SingGuard-Bench}, a multimodal guardrail benchmark with 56{,}340 examples spanning 80+ fine-grained risk types across multimodal QA, adversarial attack, and dynamic-rule evaluation settings, including cross-modal joint-risk cases where each modality is harmless in isolation but their composition implies unsafe intent. Across six benchmark families (35 datasets), SingGuard achieves state-of-the-art average F1 in every family. Dynamic-rule evaluation further shows improved policy-following accuracy from 0.6465 to 0.7415 under runtime policy shifts. Our code is available at https://github.com/inclusionAI/Sing-Guard.