In this paper, we study the harmlessness alignment problem of multimodal large language models~(MLLMs). We conduct a systematic empirical analysis of the harmlessness performance of representative MLLMs and reveal that the image input poses the alignment vulnerability of MLLMs. Inspired by this, we propose a novel jailbreak method named HADES, which hides and amplifies the harmfulness of the malicious intent within the text input, using meticulously crafted images. Experimental results show that HADES can effectively jailbreak existing MLLMs, which achieves an average Attack Success Rate~(ASR) of 90.26% for LLaVA-1.5 and 71.60% for Gemini Pro Vision. Our code and data will be publicly released.
Visual instruction tuning is an essential approach to improving the zero-shot generalization capability of Multi-modal Large Language Models (MLLMs). A surge of visual instruction datasets with various focuses and characteristics have been proposed recently, enabling MLLMs to achieve surprising results on evaluation benchmarks. To develop more capable MLLMs, in this paper, we aim to investigate a more fundamental question: ``what makes for good visual instructions?''. By conducting a comprehensive empirical study, we find that instructions focused on complex visual reasoning tasks are particularly effective in improving the performance of MLLMs on evaluation benchmarks. Building upon this finding, we design a systematic approach to automatically creating high-quality complex visual reasoning instructions. Our approach employs a synthesis-complication-reformulation paradigm, leveraging multiple stages to gradually increase the complexity of the instructions while guaranteeing quality. Based on this approach, we create the synthetic visual reasoning instruction dataset consisting of 32K examples, namely ComVint, and fine-tune four MLLMs on it. Experimental results demonstrate that our dataset consistently enhances the performance of all the compared MLLMs, e.g., improving the performance of MiniGPT-4 and BLIP-2 on MME-Cognition by 32.6% and 28.8%, respectively. Our code and data are publicly available at the link: https://github.com/RUCAIBox/ComVint.
Although pre-trained language models (PLMs) have shown impressive performance by text-only self-supervised training, they are found lack of visual semantics or commonsense, e.g., sizes, shapes, and colors of commonplace objects. Existing solutions often rely on explicit images for visual knowledge augmentation (requiring time-consuming retrieval or generation), and they also conduct the augmentation for the whole input text, without considering whether it is actually needed in specific inputs or tasks. To address these issues, we propose a novel visually-augmented fine-tuning approach that can be generally applied to various PLMs or NLP tasks, without using any retrieved or generated images, namely VAWI. Specifically, we first identify the visually-hungry words (VH-words) from input text via a token selector, where three different methods have been proposed, including syntax-, attention- and learning-based strategies. Then, we adopt a fixed CLIP text encoder to generate the visually-augmented representations of these VH-words. As it has been pre-trained by vision-language alignment task on the large-scale corpus, it is capable of injecting visual semantics into the aligned text representations. Finally, the visually-augmented features will be fused and transformed into the pre-designed visual prompts based on VH-words, which can be inserted into PLMs to enrich the visual semantics in word representations. We conduct extensive experiments on ten NLP tasks, i.e., GLUE benchmark, CommonsenseQA, CommonGen, and SNLI-VE. Experimental results show that our approach can consistently improve the performance of BERT, RoBERTa, BART, and T5 at different scales, and outperform several competitive baselines significantly. Our codes and data are publicly available at~\url{https://github.com/RUCAIBox/VAWI}.