Video Understanding on TGIF
51.4AccuracyVideo-ChatGPT
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Video-ChatGPT2026.04 | 51.4 | 3 | — | — | |
| ApETBackbone=Video-LLaVA-7B, Retained Tokens=2562026.02 | 47.6 | 3.37 | — | — | |
| Video-LLaVA2026.05 | 47.1 | — | — | — | |
| Video-LLaVA-7BBase Model=Video-LLaVA-7B2026.04 | 47 | 3.4 | — | — | |
| VanillaBackbone=Video-LLaVA-7B, Retained Tokens=20482026.02 | 46.9 | 3.34 | — | — | |
| VanillaBackbone=Video-LLaVA-7B, Original visual tokens=2048, Retained visual tokens=256, Reduction ratio=87.5%2026.05 | 46.9 | 3.34 | 100 | 100 | |
| Video-LLaVARetained Tokens=20482026.06 | 46.9 | 3.34 | — | — | |
| Video-LLaVA-7B + HalfVBase Model=Video-LLaVA-7B, Technique=HalfV2026.04 | 46.5 | 3.5 | — | — | |
| LRCPBackbone=Video-LLaVA-7B, Original visual tokens=2048, Retained visual tokens=256, Reduction ratio=87.5%2026.05 | 46.5 | 3.34 | 99.1 | 100 | |
| EvoCutRetained Tokens=2562026.06 | 46.4 | 3.34 | — | — | |
| Video-LLaVA-7B + DARTBase Model=Video-LLaVA-7B, Technique=DART2026.04 | 46.3 | 3.4 | — | — | |
| Video-LLaVA-7B + HoloVBase Model=Video-LLaVA-7B, Technique=HoloV2026.04 | 46.1 | 3.2 | — | — | |
| SparseVLMBackbone=Video-LLaVA-7B, Retained Tokens=2562026.02 | 45.9 | 3.32 | — | — | |
| SparseVLMRetained Tokens=2562026.06 | 45.9 | 3.32 | — | — | |
| Video-LLaVA-7B + FastVBase Model=Video-LLaVA-7B, Technique=FastV2026.04 | 45.2 | 3.1 | — | — | |
| OccamToken2026.05 | 44.9 | — | — | — | |
| SparseVLM2026.05 | 44.7 | — | — | — | |
| VisionZipBackbone=Video-LLaVA-7B, Retained Tokens=2562026.02 | 44.3 | 3.29 | — | — | |
| VisionZipBackbone=Video-LLaVA-7B, Original visual tokens=2048, Retained visual tokens=256, Reduction ratio=87.5%2026.05 | 44.3 | 3.29 | 94.5 | 98.5 | |
| VisionZipRetained Tokens=2562026.06 | 44.3 | 3.29 | — | — | |
| FastVBackbone=Video-LLaVA-7B, Retained Tokens=2562026.02 | 44.2 | 3.29 | — | — | |
| FastVRetained Tokens=2562026.06 | 44.2 | 3.29 | — | — | |
| VisionZip2026.05 | 42.4 | — | — | — | |
| PDropBackbone=Video-LLaVA-7B, Retained Tokens=2562026.02 | 40.3 | 3.21 | — | — | |
| PDropBackbone=Video-LLaVA-7B, Original visual tokens=2048, Retained visual tokens=256, Reduction ratio=87.5%2026.05 | 40.3 | 3.21 | 85.9 | 96.1 | |
| PDropRetained Tokens=2562026.06 | 40.3 | 3.21 | — | — | |
| VideoChat-7B2026.04 | 34.4 | 2.3 | — | — | |
| Video-LLaVARetained Tokens=2048, Reduction Ratio=0%2026.04 | 19.8 | 2.53 | — | — | |
| ID-SelectionRetained Tokens=256, Reduction Ratio=87.5%, Importance Scoring=unified score (Eq. 6)2026.04 | 17.3 | 2.46 | — | — | |
| ID-SelectionRetained Tokens=256, Reduction Ratio=87.5%, Importance Scoring=visual encoder [CLS] attention2026.04 | 17 | 2.46 | — | — | |
| FastVRetained Tokens=256, Reduction Ratio=87.5%2026.04 | 15.7 | 2.51 | — | — | |
| CDPrunerRetained Tokens=256, Reduction Ratio=87.5%2026.04 | 15.6 | 2.42 | — | — | |
| FasterVLMRetained Tokens=256, Reduction Ratio=87.5%2026.04 | 15.5 | 2.41 | — | — | |
| DivPruneRetained Tokens=256, Reduction Ratio=87.5%2026.04 | 14.8 | 2.4 | — | — | |
| CLSE-MEvaluator=Gemini-2.5-Flash, Pruning Strategy=w/ Token Merging, Retained Tokens=1942026.06 | 14.7 | 0.79 | — | — | |
| ID-SelectionRetained Tokens=128, Reduction Ratio=93.75%, Importance Scoring=visual encoder [CLS] attention2026.04 | 14.6 | 2.41 | — | — | |
| ID-SelectionRetained Tokens=128, Reduction Ratio=93.75%, Importance Scoring=unified score (Eq. 6)2026.04 | 14.4 | 2.38 | — | — | |
| VisPrunerRetained Tokens=256, Reduction Ratio=87.5%2026.04 | 14.3 | 2.39 | — | — | |
| Video-LLaVA-7BEvaluator=Gemini-2.5-Flash, Pruning Strategy=None, Retained Tokens=20482026.06 | 14.2 | 0.77 | — | — | |
| FastVRetained Tokens=128, Reduction Ratio=93.75%2026.04 | 14.1 | 2.38 | — | — | |
| FasterVLMRetained Tokens=128, Reduction Ratio=93.75%2026.04 | 13.8 | 2.36 | — | — | |
| CDPrunerRetained Tokens=128, Reduction Ratio=93.75%2026.04 | 13.8 | 2.37 | — | — | |
| DivPruneRetained Tokens=128, Reduction Ratio=93.75%2026.04 | 13.7 | 2.38 | — | — | |
| VisPrunerRetained Tokens=128, Reduction Ratio=93.75%2026.04 | 13.4 | 2.36 | — | — | |
| SparseVLMEvaluator=Gemini-2.5-Flash, Pruning Strategy=w/ Token Merging, Retained Tokens=1942026.06 | 13.1 | 0.71 | — | — | |
| CLSEEvaluator=Gemini-2.5-Flash, Pruning Strategy=Hard Pruning, Retained Tokens=1942026.06 | 12.9 | 0.69 | — | — | |
| Video-LLaVA-7BEvaluator=GPT-4o-mini, Pruning Strategy=None, Retained Tokens=20482026.06 | 11.3 | 1.19 | — | — | |
| CLSE-MEvaluator=GPT-4o-mini, Pruning Strategy=w/ Token Merging, Retained Tokens=1942026.06 | 11.2 | 1.05 | — | — | |
| SparseVLMEvaluator=GPT-4o-mini, Pruning Strategy=w/ Token Merging, Retained Tokens=1942026.06 | 10.7 | 1.03 | — | — | |
| CLSEEvaluator=GPT-4o-mini, Pruning Strategy=Hard Pruning, Retained Tokens=1942026.06 | 9.6 | 1.06 | — | — | |
| FiCoCo-VEvaluator=Gemini-2.5-Flash, Pruning Strategy=Hard Pruning, Retained Tokens=1942026.06 | 6.7 | 0.38 | — | — | |
| PDropEvaluator=Gemini-2.5-Flash, Pruning Strategy=Hard Pruning, Retained Tokens=1942026.06 | 6.1 | 0.37 | — | — | |
| FiCoCo-VEvaluator=GPT-4o-mini, Pruning Strategy=Hard Pruning, Retained Tokens=1942026.06 | 5.6 | 1.18 | — | — | |
| PDropEvaluator=GPT-4o-mini, Pruning Strategy=Hard Pruning, Retained Tokens=1942026.06 | 5 | 0.96 | — | — | |
| FastVEvaluator=GPT-4o-mini, Pruning Strategy=Hard Pruning, Retained Tokens=1942026.06 | 2.8 | 1.16 | — | — | |
| FastVEvaluator=Gemini-2.5-Flash, Pruning Strategy=Hard Pruning, Retained Tokens=1942026.06 | 2.3 | 0.14 | — | — |