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

AI systems need to know when to say "I don't know": The paper identifies a critical gap in current Visual Question Answering models—they lack the ability to appropriately abstain from answering questions they cannot reliably answer, which is essential for building trustworthy AI systems that users can depend on.

Current multi-modal AI models perform poorly on unanswerable questions: When tested on the UNK-VQA dataset, several state-of-the-art multi-modal large models showed "significant limitations" in recognizing when questions cannot be answered, suggesting these widely-used systems may confidently provide incorrect answers rather than admitting uncertainty.

The challenge goes beyond simple image replacement: The dataset deliberately creates difficult cases where questions remain semantically close to answerable ones but become unanswerable through subtle perturbations—this is more realistic and challenging than simply swapping out images, making it a more rigorous test of genuine understanding versus pattern matching.

Benchmarking tools are needed to improve AI reliability: By providing a comprehensive dataset specifically designed to test abstention ability, the researchers create a measurable standard that can help developers identify and fix this weakness in their models before deployment in real-world applications.

Responsible AI development requires targeted evaluation: The paper demonstrates that improving trustworthiness requires not just building better models, but creating specific evaluation benchmarks that probe for particular failure modes—in this case, the tendency to answer questions the model shouldn't attempt.