Datasets:
license: cc-by-nc-4.0
language:
- en
pretty_name: MIPBench
size_categories:
- 1K<n<10K
task_categories:
- visual-question-answering
tags:
- vision-language-models
- multi-image
- benchmark
- evaluation
- position-bias
- visual-question-answering
MIPBench
MIPBench is an evaluation-only benchmark for measuring position sensitivity in position-invariant multi-image visual question answering. Each example contains multiple images, a question, an answer, and provenance metadata. The benchmark is intended to evaluate whether a vision-language model changes its answer when the input images are permuted while the ground-truth answer remains unchanged.
Dataset Summary
MIPBench contains 6,994 position-invariant multi-image VQA examples derived from nine public source benchmarks. The examples are filtered by a three-stage pipeline and manual review to remove questions whose correct answer depends on image order.
The benchmark is organized into five task categories and three answer formats. It is designed for evaluating accuracy, consistency rate, and flip rate under image permutations.
Intended Use
MIPBench is intended for controlled evaluation of position sensitivity in multi-image vision-language models. It can be used to test whether model predictions remain stable when input images are reordered in tasks where the ground-truth answer should remain unchanged.
Out-of-Scope Use
MIPBench is not intended for model training, general-purpose VLM ranking, or deployment qualification in high-stakes domains such as medicine, transportation, law, or safety-critical decision making.
Data Fields
id: MIPBench example identifier.question_type: answer format, such as multiple choice, yes/no, or short answer.question: the multi-image question.choices: answer options for multiple-choice examples, otherwise null.answer: ground-truth answer.images: list of input images.source: source benchmark.class: MIPBench taxonomy category.raw_id: original source-benchmark example identifier.
Provenance
MIPBench is derived from nine public multi-image benchmarks. Each retained example preserves its source benchmark and original identifier through the source and raw_id fields.
Limitations
MIPBench is a high-precision benchmark for position-invariant multi-image VQA, not a comprehensive benchmark of general multi-image reasoning. Because it is derived from existing public benchmarks, it may inherit their linguistic, geographic, demographic, visual-domain, and licensing biases. Questions are English-only. Some residual ambiguity in position invariance may remain despite filtering and manual review.