← All benchmarks

MMAR

Multimodal · 2025-05-19

MMAR is a 1,000-item audio reasoning benchmark of audio-question-answer triplets from real-world videos spanning sound, music, speech and their mixtures, organized into signal, perception, semantic and cultural reasoning layers. Questions are multiple choice with mostly four (sometimes two or three) options, and scores are accuracy.

Top models (higher is better)