Hallucination benchmarks for multimodal large language models usually test whether a model can pick the correct answer from a set of distractors. That setup ignores a key failure mode: confidently choosing one of several wrong options when none is supported by the image. To address this, the authors introduce HumbleBench, a benchmark built from the Panoptic Scene Graph dataset that always includes a "None of the above" (NOTA) option alongside object, relation, and attribute questions. GPT-4-Turbo generated the questions, and human verification produced 22,831 items.
The study evaluates 27 models, spanning general-purpose, reasoning-oriented, and proprietary systems. Using accuracy plus NOTA-aware metrics, diagnostic stress tests, and a human comparison, the authors find that even the strongest model selects NOTA in only about half of the questions where that answer is correct. Adding cautious prompting shifts model behavior toward choosing NOTA more often, but it does not improve their underlying ability to detect when an option is unsupported. Human participants rejected unsupported options considerably more reliably than every model tested.
The results suggest that standard recognition accuracy paints an incomplete picture of hallucination robustness. A model can score well on forced-choice benchmarks while still failing to admit that an image does not support any given answer. The authors argue that false-option rejection is a directly observable facet of epistemic humility, and they release their code and dataset publicly for further research.