Video Understanding on VCRBench
33.8AccuracyVecAtten.
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| VecAtten.Backbone=InternVL-3.5-8B, #Frames=64, #Tokens=17K, Average Sparsity=78.62026.03 | 33.8 | — | |
| Full Atten.Backbone=InternVL-3.5-8B, #Frames=64, #Tokens=17K, Average Sparsity=0.02026.03 | 32.9 | — | |
| XAtten.Backbone=InternVL-3.5-8B, #Frames=64, #Tokens=17K, Average Sparsity=78.12026.03 | 32.5 | — | |
| AnchorAtten.Backbone=InternVL-3.5-8B, #Frames=64, #Tokens=17K, Average Sparsity=78.62026.03 | 31.3 | — | |
| FlexPref.Backbone=InternVL-3.5-8B, #Frames=64, #Tokens=17K, Average Sparsity=76.52026.03 | 30 | — | |
| Full Atten.Backbone=Qwen2.5-VL-7B-Instruct, #FPS=1, #Tokens=26K, Average Sparsity=0.02026.03 | 25.8 | — | |
| VecAtten.Backbone=Qwen2.5-VL-7B-Instruct, #FPS=1, #Tokens=26K, Average Sparsity=78.52026.03 | 25.4 | — | |
| AnchorAtten.Backbone=Qwen2.5-VL-7B-Instruct, #FPS=1, #Tokens=26K, Average Sparsity=74.62026.03 | 22.9 | — | |
| FlexPref.Backbone=Qwen2.5-VL-7B-Instruct, #FPS=1, #Tokens=26K, Average Sparsity=73.62026.03 | 22.5 | — | |
| XAtten.Backbone=Qwen2.5-VL-7B-Instruct, #FPS=1, #Tokens=26K, Average Sparsity=73.62026.03 | 20 | — | |
| Gemini 2.5 Pro2026.03 | — | 53.4 | |
| Gemini-3-Pro2026.03 | — | 51.4 | |
| Seed-1.5-VL2026.03 | — | 51.8 | |
| Seed1.82026.03 | — | 59.8 |