HallusionBench (OpenVLM)
Multimodal
HallusionBench probes visual illusion and hallucination with paired yes/no questions. The leaderboard's headline figure combines the question-level, figure-level, and all-question accuracies that it also reports separately, so it is a composite rather than a single item proportion. Scored by the OpenVLM Leaderboard, which runs every model through OpenCompass's VLMEvalKit under one fixed harness. That is a different measurement protocol from the lab-reported and Artificial Analysis runs of the same underlying benchmarks, so it forms its own factor group. No evaluation_item_count is declared: the leaderboard publishes scores to one decimal place, which is coarser than the item lattice, so the denominator cannot be confirmed from the numbers and a binomial noise model is not claimed.