The essence of the matter lies in the ability of Vision Language Models (VLMs) to discern tasks they cannot solve, an ability akin to a human recognizing questions that are unanswerable based on the information at hand. This concept, termed Unsolvable Problem Detection (UPD), evaluates whether VLMs can identify and abstain from responding to questions that do not align with corresponding visual cues or are inherently unsolvable. It is a new benchmark for testing the limits of AI comprehension and honesty in addressing its own limitations. In recent history, evaluations of VLMs have traditionally revolved around their proficiency in generating correct responses to solvable problems. UPD, however, shifts this narrative by introducing challenges that require VLMs to express restraint rather than just knowledge, an aspect that brings AI closer to nuanced human thought processes. This evolution in assessment strategies acknowledges the growing complexity of questions face and the necessity for models to recognize the bounds of their capabilities. Unsolvable problems, as defined in recent research, are classified into three types: Absent Answer Detection (AAD), Incompatible Answer Set Detection (IASD), and Incompatible Visual Question Detection (IVQD). AAD involves cases where the correct answer is not presented among the options, IASD where none of the provided answers relate to the visual content, and IVQD where the question is irrelevant to the accompanying image. These categories aim to test the discernment abilities of VLMs in varied contexts. Evaluations using the MMBench dataset have shown that VLMs, even advanced ones like GPT-4V and LLaVA-NeXT, struggle with UPD tasks despite their proficiency with standard questions. This reveals a significant hurdle in the quest for AI models that can reliably navigate both conventional and unsolvable queries. Nevertheless, larger models tend to fare better than their smaller counterparts, suggesting model capacity plays a role in UPD performance. The study explored whether prompt engineering, including the addition of options like “None of the above” and specific instructions, could help VLMs better tackle UPD. The results were mixed, with different models responding variably to these strategies. Instruction tuning, a method involving training adjustments, showed promise, but challenges persist, especially with smaller VLMs and AAD tasks. These findings underscore the complex nature of UPD and the need for continued innovation. A scientific paper titled “End-to-end Evaluation of Visual Question Answering Models under a Resource Constrained Environment,” published in the Journal of
Artificial intelligence Research, delves into the broader context of VLM evaluation. It examines VLMs in resource-constrained situations, an environment that could similarly accentuate the need for models to recognize their limitations, as in UPD tasks. This correlation suggests research in UPD could have broader implications for the deployment of AI in varying operational conditions. The findings from UPD research paint a picture of a future where VLMs not only answer questions but also understand their own limits, a crucial step towards AI that resembles human-like discernment. This understanding is essential for building trust in AI systems and ensuring their decisions are reliable. As the field progresses, it is anticipated that models will not only grow in knowledge but also in the wisdom to know when to apply it—or importantly, when not to.