GuardAlign: Test-time Safety Alignment in Multimodal Large Language Models

TL;DR

GuardAlign is a training-free safety framework for multimodal large language models that improves detection of unsafe content and maintains safety signals during output generation, enhancing safety without sacrificing performance.

Contribution

It introduces OT-enhanced safety detection and cross-modal attentive calibration, providing an effective, training-free safety mechanism for multimodal models.

Findings

Reduces unsafe responses by up to 39% on SPA-VL

Improves VQAv2 accuracy from 78.51% to 79.21%

Maintains model utility while enhancing safety

Abstract

Large vision-language models (LVLMs) have achieved remarkable progress in vision-language reasoning tasks, yet ensuring their safety remains a critical challenge. Recent input-side defenses detect unsafe images with CLIP and prepend safety prefixes to prompts, but they still suffer from inaccurate detection in complex scenes and unstable safety signals during decoding. To address these issues, we propose GuardAlign, a training-free defense framework that integrates two strategies. First, OT-enhanced safety detection leverages optimal transport to measure distribution distances between image patches and unsafe semantics, enabling accurate identification of malicious regions without additional computational cost. Second, cross-modal attentive calibration strengthens the influence of safety prefixes by adaptively reallocating attention across layers, ensuring that safety signals remain consistently activated throughout generation. Extensive evaluations on six representative MLLMs demonstrate that GuardAlign reduces unsafe response rates by up to 39% on SPA-VL, while preserving utility, achieving an improvement on VQAv2 from 78.51% to 79.21%.