MMIU

A 11,698-question multi-image multiple-choice benchmark spanning 52 tasks and 7 image-relationship types, scored by accuracy.

Also known as: Multimodal Multi-image Understanding, MMIU-Benchmark

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categorymultimodal
Subcategorymulti-image multiple-choice understanding
Page statusactive
Metricaccuracy
Directionhigher_is_better
Unit%
Dataset size11698
Dataset licenceCC-BY-4.0
PublisherOpenGVLab (Shanghai AI Laboratory and collaborators)

What it measures

MMIU tests whether a vision-language model can answer a multiple-choice question about several images at once. Items cover seven relationship types, from low-level forensic detection through semantic correspondence, temporal ordering, and 2D/3D spatial tasks. Each sample is a question, an option list, a list of image paths, and a letter answer. Table 1 of arXiv:2408.02718 names five modalities (image, text, video, point cloud, depth). Language of the questions is English.

Task format

Multiple-choice with images. inspect_evals builds a user message from the question field plus the image list, uses Inspect's multiple_choice solver, and scores with choice() plus per-task accuracy. Shuffle defaults to true. A task_name filter can run one of the inspect task keys. Authors recommend VLMEvalKit as another runner.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub