POPE Hallucination Evaluation
Polling-based Object Probing Evaluation â€" binary yes/no answers across three difficulty categories. Lower hallucination rate = better. Dataset: MSCOCO val2014 images.
No VP runs found.
Launch evaluation â†'Methodology Note
POPE asks "Is there a [object] in the image?" â€" a binary yes/no question for each image. Hallucination rate = fraction of questions where the model affirms a non-existent object. Three difficulty levels: random (easiest), popular (objects often co-occurring),adversarial (objects most likely hallucinated). Measured pre- and post-MITCF intervention.