Jianyu Liu

Jianyu Liu, Hangyu Guo, Ranjie Duan et al.

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 \textbf{DREAM} (\textit{\textbf{D}isentangling \textbf{R}isks to \textbf{E}nhance Safety \textbf{A}lignment in \textbf{M}LLMs}), 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. The data and code are available at https://github.com/Kizna1ver/DREAM.

Jianyu Liu, Hegler Tissot

Learning knowledge representation is an increasingly important technology applicable in many domain-specific machine learning problems. We discuss the effectiveness of traditional Link Prediction or Knowledge Graph Completion evaluation protocol when embedding knowledge representation for categorised multi-relational data in the clinical domain. Link prediction uses to split the data into training and evaluation subsets, leading to loss of information along training and harming the knowledge representation model accuracy. We propose a Clustering Evaluation Protocol as a replacement alternative to the traditionally used evaluation tasks. We used embedding models trained by a knowledge embedding approach which has been evaluated with clinical datasets. Experimental results with Pearson and Spearman correlations show strong evidence that the novel proposed evaluation protocol is pottentially able to replace link prediction.