VisuoAlign: Safety Alignment of LVLMs with Multimodal Tree Search
MingSheng Li, Guangze Zhao, Sichen Liu
TL;DR
VisuoAlign tackles safety alignment for LVLMs in multimodal settings where visual inputs introduce cross-modal jailbreak risks and reasoning safety can be under-monitored. It introduces prompt-guided tree search to embed safety constraints into multimodal reasoning and uses Monte Carlo Tree Search to generate a diverse, adversarial safety-alignment dataset. Dynamic safety scaling during inference prunes unsafe reasoning paths and yields compliant outputs. Across four benchmarks and multiple models, VisuoAlign achieves state-of-the-art robustness, with Jailbreak Success Rate reductions exceeding 57% on average and Alignment Safety Rate above 0.92, while demonstrating solid cross-model generalization and practical efficiency.
Abstract
Large Vision-Language Models (LVLMs) have achieved remarkable progress in multimodal perception and generation, yet their safety alignment remains a critical challenge.Existing defenses and vulnerable to multimodal jailbreaks, as visual inputs introduce new attack surfaces, reasoning chains lack safety supervision, and alignment often degrades under modality fusion.To overcome these limitation, we propose VisuoAlign, a framework for multi-modal safety alignment via prompt-guided tree search.VisuoAlign embeds safety constrains into the reasoning process through visual-textual interactive prompts, employs Monte Carlo Tree Search(MCTS) to systematically construct diverse safety-critical prompt trajectories, and introduces prompt-based scaling to ensure real-time risk detection and compliant responses.Extensive experiments demonstrate that VisuoAlign proactively exposes risks, enables comprehensive dataset generation, and significantly improves the robustness of LVLMs against complex cross-modal threats.