Measuring Epistemic Humility in Multimodal Large Language Models

MLLMs struggle with admitting uncertainty: Even the strongest multimodal AI models tested correctly rejected false options only about 50% of the time, revealing a significant gap in their ability to recognize when they lack sufficient information to answer a question—a critical capability for trustworthy AI systems.

Current AI benchmarks miss an important safety measure: Existing hallucination tests focus on whether models can pick the right answer from options, but overlook whether they can recognize when none of the options are correct. HumbleBench's inclusion of "None of the above" options revealed this blind spot in how we evaluate AI safety.

Telling models to be cautious doesn't solve the core problem: When researchers prompted models to be more cautious, the models simply chose "None of the above" more often—but without actually improving their ability to detect when options were genuinely unsupported by the image, suggesting a superficial rather than genuine improvement in judgment.

The benchmark covers real-world hallucination types: HumbleBench evaluated 22,831 questions testing three types of hallucinations (object, relation, and attribute), providing a comprehensive assessment of where multimodal models fail across different categories of visual understanding.

Humans outperform AI at recognizing limits: The study found that humans reject unsupported options considerably more often than current AI models, indicating that developing this form of "epistemic humility"—knowing what you don't know—remains an important challenge for making AI systems more trustworthy.