OmniCompass: A Holistic Benchmark for Omni-Modal Video Grounding and Understanding
New benchmark for omni-modal video grounding with a vision-language track (820 videos, 2,460 instances) and an audio-visual track (240 videos, 720 instances), each tested under paired text-only vs text+image settings. Evaluation of 50 multimodal LLMs shows open-source models trailing proprietary ones, especially on audio-visual understanding.
Why it matters
Top upvoted paper of the window with 566 upvotes on HuggingFace Daily Papers for Oct 16, 2026.
Importance: 4/5
flagship/notable release with independent signal (upvotes or multi-source coverage)
Sources
official
arXiv:2610.18829 — OmniCompass