Long Video Understanding on CG-Bench long
38.6Accuracy (long-acc)InternVL3-8B
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| InternVL3-8BNum Frame=-, Token/Frame=-, Total Num of Token=>12k2026.04 | 38.6 | — | |
| POINTS1.5-8B (baseline)Num Frame=64, Token/Frame=324, Total Num of Token=≈ 20K2026.04 | 36.7 | 53.9 | |
| POINTS-Long (standby)Num Frame=64, Token/Frame=32, Total Num of Token=2048 (10%)2026.04 | 35.7 | 53.8 | |
| POINTS-Long (focus)Num Frame=64, Token/Frame=324+32, Total Num of Token=≈ 22K2026.04 | 35.4 | 54.4 | |
| POINTS-Long (standby)Num Frame=64, Token/Frame=16, Total Num of Token=1024 (5%)2026.04 | 34.6 | 52.9 | |
| POINTS-Long (standby)Num Frame=64, Token/Frame=8, Total Num of Token=512 (2.5%)2026.04 | 33.6 | 52.7 | |
| LLaVA-OV-7BNum Frame=64, Token/Frame=-, Total Num of Token=>12k2026.04 | 30.9 | — |