UNK-VQA: A Dataset and a Probe into the Abstention Ability of Multi-modal Large Models

TL;DR

UNK-VQA tackles abstention in visual question answering by building a perturbation-based dataset that preserves semantic coherence while producing unanswerable cases. It combines five perturbation types, human annotations, and a simple selective-classifier approach to enable models to abstain, while revealing robustness gaps in contemporary multi-modal large models through zero- and few-shot evaluations and LLaVA fine-tuning. The work demonstrates that current models struggle to reliably abstain, but gains are achievable via targeted prompting, post-hint explanations, and supervised fine-tuning on multi-modal bases. Together, the dataset and method provide a practical benchmark and baseline toward more trustworthy VQA systems and point to directions for future multi-modal learning and evaluation.

Abstract

Teaching Visual Question Answering (VQA) models to refrain from answering unanswerable questions is necessary for building a trustworthy AI system. Existing studies, though have explored various aspects of VQA but somewhat ignored this particular attribute. This paper aims to bridge the research gap by contributing a comprehensive dataset, called UNK-VQA. The dataset is specifically designed to address the challenge of questions that models do not know. To this end, we first augment the existing data via deliberate perturbations on either the image or question. In specific, we carefully ensure that the question-image semantics remain close to the original unperturbed distribution. By this means, the identification of unanswerable questions becomes challenging, setting our dataset apart from others that involve mere image replacement. We then extensively evaluate the zero- and few-shot performance of several emerging multi-modal large models and discover their significant limitations when applied to our dataset. Additionally, we also propose a straightforward method to tackle these unanswerable questions. This dataset, we believe, will serve as a valuable benchmark for enhancing the abstention capability of VQA models, thereby leading to increased trustworthiness of AI systems. We have made the dataset (https://github.com/guoyang9/UNK-VQA) available to facilitate further exploration in this area.

Figures (9)

• Figure 1: Left subfigure: Five perturbation types from our UNK-VQA dataset and their corresponding exemplars. Right subfigure: The unanswerability ratio with respect to each perturbation type.
• Figure 2: Illustration of text-based perturbations. We utilize a strong multi-modal model BLIP blip to generate answers (Base-A) to the perturbed question (Pert-Q). The left example shows that when we replace the anchor noun word with an alternative, the baseline still generates a reasonable answer, even though the modified question becomes unanswerable. Regarding the semantic negation examples, each question can have an uncountable number of potential answers.
• Figure 3: Illustration of three image-based perturbations. The BLIP model is also employed to generate answers for the perturbed image (Base-A). In the case of image replacement samples, we replace the original image with another image that shares a high degree of semantic similarity. For the latter two perturbation types, we cover the most relevant object with a mask and other regions of this image, respectively.
• Figure 4: Sunburst distribution of the first four words in the UNK-VQA dataset questions. Most questions begin with the word 'what'.
• Figure 5: Labeling confidence and consensus among three annotators (y-axis: counted numbers). The majority of annotators exhibits a higher level of confidence and can reach a consensus regarding the question answerability.