Dataset Card for REST (Render-Equivalence Stress Tests)

REST and REST+ measure whether Vision-Language Models (VLMs) give the same answer to the same question when the question is presented as text versus as an image of that text. Questions are drawn from established text benchmarks and rendered in clean, readable form, so that OCR stays easy and inconsistencies reflect reasoning across modalities rather than failures to read.

• Paper: Same content, different answers: Cross-modal inconsistency in mllms

Dataset Summary

Each question is available in three input conditions:

Condition	Description
text	Standard textual prompt.
image	The full prompt rendered as an image, requiring OCR.
partial	Context (or the question, for multiple choice) as an image; the remaining options or final question as text.

REST renders each question once (DejaVu Sans, black text on white background, 200 DPI). REST+ adds 10 visual permutations per question: 3 font families (DejaVu Sans, Courier New, a cursive font) × 3 resolutions (50, 100, 200 DPI), plus a colored-text variant. REST+ uses a 10% sample of MMLU.