OmniGuard: Unified Omni-Modal Guardrails with Deliberate Reasoning

2025-12-02

AI Summary

Existing binary-classification-based unimodal safety mitigation methods exhibit insufficient robustness for Omni-modal Large Language Models (OLLMs) facing security alignment challenges across diverse modalities—text, image, video, and audio. Method: We propose the first omni-modal safety framework enabling deliberative reasoning. It introduces a cross-modal unified safety representation system, incorporating structured safety labels and expert-model-distilled safety annotations to enable fine-grained, interpretable risk identification. The framework integrates multimodal classification, cross-modal alignment, and targeted knowledge distillation, trained on a large-scale omni-modal safety dataset. Contribution/Results: Extensive evaluation across 15 benchmarks demonstrates significant improvements in safety mitigation’s unity, generalizability, and reliability—marking the first framework to holistically address omni-modal safety alignment with interpretable, reasoning-aware mechanisms.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Safety and Robustness
• Computer Vision: Multi-modal Vision

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Security and Privacy: Large-scale security measurements

Abstract

Omni-modal Large Language Models (OLLMs) that process text, images, videos, and audio introduce new challenges for safety and value guardrails in human-AI interaction. Prior guardrail research largely targets unimodal settings and typically frames safeguarding as binary classification, which limits robustness across diverse modalities and tasks. To address this gap, we propose OmniGuard, the first family of omni-modal guardrails that performs safeguarding across all modalities with deliberate reasoning ability. To support the training of OMNIGUARD, we curate a large, comprehensive omni-modal safety dataset comprising over 210K diverse samples, with inputs that cover all modalities through both unimodal and cross-modal samples. Each sample is annotated with structured safety labels and carefully curated safety critiques from expert models through targeted distillation. Extensive experiments on 15 benchmarks show that OmniGuard achieves strong effectiveness and generalization across a wide range of multimodal safety scenarios. Importantly, OmniGuard provides a unified framework that enforces policies and mitigates risks in omni-modalities, paving the way toward building more robust and capable omnimodal safeguarding systems.

Problem

• Develops omni-modal safety guardrails for multimodal AI
• Addresses limitations of unimodal binary classification approaches
• Enables robust safeguarding across diverse modalities and tasks

Innovation

• Unified omni-modal guardrails with deliberate reasoning
• Large curated dataset with structured safety labels
• Generalization across diverse multimodal safety scenarios