Multimodal Large Language Models (MLLMs) have expanded the capabilities of traditional language models by enabling interaction through both text and images. However, ensuring the safety of these models remains a significant challenge, particularly in accurately identifying whether multimodal content is safe or unsafe—a capability we term safety awareness. In this paper, we introduce MMSafeAware, the first comprehensive multimodal safety awareness benchmark designed to evaluate MLLMs across 29 safety scenarios with 1,500 carefully curated image-prompt pairs. MMSafeAware includes both unsafe and over-safety subsets to assess models’ abilities to correctly identify unsafe content and avoid over-sensitivity that can hinder helpfulness. Evaluating nine widely used MLLMs using MMSafeAware reveals that current models are not sufficiently safe and often overly sensitive; for example, GPT-4V misclassifies 36.1% of unsafe inputs as safe and 59.9% of benign inputs as unsafe. We further explore three methods to improve safety awareness—prompting-based approaches, visual contrastive decoding, and vision-centric reasoning fine-tuning—but find that none achieve satisfactory performance. Our findings highlight the profound challenges in developing MLLMs with robust safety awareness, underscoring the need for further research in this area. All the code and data will be publicly available to facilitate future research.

This paper explores the problem of commonsense level vision-knowledge conflict in Multimodal Large Language Models (MLLMs), where visual information contradicts model’s internal commonsense knowledge. To study this issue, we introduce an automated framework, augmented with human-in-the-loop quality control, to generate inputs designed to simulate and evaluate these conflicts in MLLMs. Using this framework, we have crafted a diagnostic benchmark consisting of 374 original images and 1,122 high-quality question-answer (QA) pairs. The benchmark covers two aspects of conflict and three question types, providing a thorough assessment tool. We apply this benchmark to assess the conflict-resolution capabilities of nine representative MLLMs from various model families. Our results indicate an evident over-reliance on parametric knowledge for approximately 20% of all queries, especially among Yes-No and action-related problems. Based on these findings, we evaluate the effectiveness of existing approaches to mitigating the conflicts and compare them to our “Focus-on-Vision” prompting strategy. Despite some improvement, the vision-knowledge conflict remains unresolved and can be further scaled through our data construction framework. Our proposed framework, benchmark, and analysis contribute to the understanding and mitigation of vision-knowledge conflicts in MLLMs.

Scene Graph Generation (SGG) provides basic language representation of visual scenes, requiring models to grasp complex and diverse semantics between objects. This complexity and diversity in SGG leads to underrepresentation, where parts of triplet labels are rare or even unseen during training, resulting in imprecise predictions. To tackle this, we propose integrating the pretrained Vision-language Models to enhance representation. However, due to the gap between pretraining and SGG, direct inference of pretrained VLMs on SGG leads to severe bias, which stems from the imbalanced predicates distribution in the pretraining language set. To alleviate the bias, we introduce a novel LM Estimation to approximate the unattainable predicates distribution. Finally, we ensemble the debiased VLMs with SGG models to enhance the representation, where we design a certainty-aware indicator to score each sample and dynamically adjust the ensemble weights. Our training-free method effectively addresses the predicates bias in pretrained VLMs, enhances SGG’s representation, and significantly improve the performance.

Meta-learning has emerged as an effective approach for few-shot text classification. However, current studies fail to realize the importance of the semantic interaction between sentence features and neglect to enhance the generalization ability of the model to new tasks. In this paper, we integrate an adversarial network architecture into the meta-learning system and leverage cost-effective modules to build a novel few-shot classification framework named SaAML. Significantly, our approach can exploit the temporal convolutional network to encourage more discriminative representation learning and explore the attention mechanism to promote more comprehensive feature expression, thus resulting in better adaptation for new classes. Through a series of experiments on four benchmark datasets, we demonstrate that our new framework acquires considerable superiority over state-of-the-art methods in all datasets, increasing the performance of 1-shot classification and 5-shot classification by 7.15% and 2.89%, respectively.

Although pretrained Transformers such as BERT achieve high accuracy on in-distribution examples, do they generalize to new distributions? We systematically measure out-of-distribution (OOD) generalization for seven NLP datasets by constructing a new robustness benchmark with realistic distribution shifts. We measure the generalization of previous models including bag-of-words models, ConvNets, and LSTMs, and we show that pretrained Transformers’ performance declines are substantially smaller. Pretrained transformers are also more effective at detecting anomalous or OOD examples, while many previous models are frequently worse than chance. We examine which factors affect robustness, finding that larger models are not necessarily more robust, distillation can be harmful, and more diverse pretraining data can enhance robustness. Finally, we show where future work can improve OOD robustness.