The fast-paced evolution of artificial intelligence, particularly Vision Language Models (VLMs), presents challenges in ensuring their reliability and trustworthiness. These VLMs integrate visual and textual understanding, however, their increasing sophistication has brought into focus their ability to detect and not respond to unsolvable or irrelevant questions— an aspect known as Unsolvable Problem Detection (UPD).
UPD is pivotal as it teaches the model to identify situations where a question doesn’t match with the given image or lacks an appropriate answer, thus increasing trust in the model. To evaluate VLM capabilities in UPD, researchers have suggested three unique problem types: Absent Answer Detection (AAD), where the correct answer is not available; Incompatible Answer Set Detection (IASD) to test if the model can identify irrelevant answer sets; and Incompatible Visual Question Detection (IVQD) that assesses the model’s comprehension of the correlation between visual content and text-based queries.
Researchers devised benchmarks from the MMBench dataset, setting specific standards for AAD, IASD and IVQD which were then used to examine a range of advanced VLMs such as LLaVA-1.5-13B, Qwen-VL-Chat, LLaVA-NeXT, and GPT-4V (vision). The findings showcased that most VLMs find it challenging to withhold from unsolvable problems, even though their regular question-answering ability is satisfactory. Larger models like GPT-4V and LLaVA-Next-34B performed comparably better, but displayed limitations in specific areas, such as object localization and attribute comparison.
The study investigated prompt engineering strategies to improve VLM capability for UPD, including incorporating additional options or instructions. However, these methods had varying levels of impact on different VLMs, and while such modifications increased UPD accuracy, they often lessened standard accuracy. Instruction tuning, a training approach, proved to be more effective than prompt engineering in diverse settings.
The research underscored the complexity of UPD and the need for innovative techniques to enhance VLMs’ trustworthiness. Although progress has been made, it is just the beginning. Future work might include the development of reasoning chains, extending to expert-level queries, and the devising of post-detection methods.
