Multimodal Large Language Models (MLLMs) pose unique safety challenges due to their integration of visual and textual data, thereby introducing new dimensions of potential attacks and complex risk combinations. In this paper, we begin with a detailed analysis aimed at disentangling risks through step-by-step reasoning within multimodal inputs. We find that systematic multimodal risk disentanglement substantially enhances the risk awareness of MLLMs. Via leveraging the strong discriminative abilities of multimodal risk disentanglement, we further introduce DREAM ( Disentangling Risks to Enhance Safety Alignment in MLLMs), a novel approach that enhances safety alignment in MLLMs through supervised fine-tuning and iterative Reinforcement Learning from AI Feedback (RLAIF). Experimental results show that DREAM significantly boosts safety during both inference and training phases without compromising performance on normal tasks (namely oversafety), achieving a 16.17% improvement in the SIUO safe&effective score compared to GPT-4V.

Recent work has revealed that in-context learning for large language models exhibits compositional generalization capacity, which can be enhanced by selecting in-context demonstrations similar to test cases to provide contextual information. However, how to exhibit in-context compositional generalization (ICCG) of large vision-language models (LVLMs) is non-trival. Due to the inherent asymmetry between visual and linguistic modalities, ICCG in LVLMs faces an inevitable challenge—redundant information on the visual modality. The redundant information affects in-context learning from two aspects: (1) Similarity calculation may be dominated by redundant information, resulting in sub-optimal demonstration selection. (2) Redundant information in in-context demonstrations brings misleading contextual information to in-context learning. To alleviate these problems, we propose a demonstration selection method to achieve ICCG for LVLMs, by considering two key factors of demonstrations: content and structure, from a multimodal perspective. Specifically, we design a diversity-coverage-based matching score to select demonstrations with maximum coverage, and avoid selecting demonstrations with redundant information via their content redundancy and structural complexity. We build a GQA-ICCG dataset to simulate the ICCG setting, and conduct experiments on GQA-ICCG and the VQA v2 dataset. Experimental results demonstrate the effectiveness of our method.