| Rank | Model | Score |
|---|---|---|
| 1 | lfm3-omni-2b | 70.09 |
1 phaseActive
A 684-video, 1,197-question multiple-choice Audio-Visual QA benchmark requiring explicit cross-modal temporal alignment across 6 task families, evaluated on 24 foundation models under 37 model-modality settings.
Quick answer: Daily-Omni ("Towards Audio-Visual Reasoning with Temporal Alignment across Modalities") is a benchmark introduced by Ziwei Zhou, Rui Wang, Zuxuan Wu, and Yu-Gang Jiang (Fudan University, submitted May 2025, revised March 2026) with 684 real-world videos and 1,197 multiple-choice questions spanning 6 task families that explicitly require cross-modal temporal reasoning.
Most multimodal benchmarks evaluate visual and audio understanding independently. Daily-Omni instead requires models to synchronize cross-modal information over time — questions are only answerable by aligning what is seen with what is heard at the correct moment, not by either modality alone. The authors evaluated 24 foundation models under 37 model-modality settings (Audio+Video / Audio-only / Video-only / Text-only) and found many end-to-end MLLMs still struggle on alignment-critical questions.
| Model | Score | Source | Date |
|---|---|---|---|
| LFM3-Omni-2B | 70.09% | Liquid AI model card | 2026-10-27 |