Long-form audio-video understanding on WorldSense up to 10min (Avg.)
52.6Avg Performance (WorldSense)AVOC
Evaluation Results
| Method | Links | |
|---|---|---|
| AVOCLLM Backbone Scale=7-8B, Global token budget (K)=10240, Modality token ratio (Kvideo:Kaudio)=2:12026.06 | 52.6 | |
| video-SALMONN 2+ 7BLLM Backbone Scale=7-8B2026.06 | 50.9 | |
| MiniCPM-o 4.5LLM Backbone Scale=7-8B2026.06 | 50.3 | |
| HumanOmni-7BLLM Backbone Scale=7-8B2026.06 | 47.1 | |
| OmniZipLLM Backbone Scale=7-8B2026.06 | 46.3 | |
| Qwen2.5-Omni-7BLLM Backbone Scale=7-8B2026.06 | 45.4 | |
| MiniCPM-o 2.6LLM Backbone Scale=7-8B2026.06 | 44.3 | |
| VideoLLaMA2-7BLLM Backbone Scale=7-8B2026.06 | 25.4 |