Long Video Understanding on Video-MME (full)
66.4Overall PerformanceQwen2.5-VL-128
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Qwen2.5-VL-128Input Frames=128, Token Retain Ratio (r)=100% (Full Performance)2026.05 | 66.4 | 77.4 | 65.8 | 56.1 | |
| ST-SimDiff-128Input Frames=128, Token Retain Ratio (r)=50%2026.05 | 66.3 | 76.8 | 66.3 | 55.8 | |
| ST-SimDiff-128Input Frames=128, Token Retain Ratio (r)=30%2026.05 | 65.1 | 75.4 | 64.7 | 55.2 | |
| CoRDSCase=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=6K, Decoding Budget=6K2026.05 | 64.84 | 75.26 | 65.44 | 57.56 | |
| CoRDSCase=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=12K, Decoding Budget=12K2026.05 | 64.84 | 75.69 | 65.66 | 56.9 | |
| CoRDSCase=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=24K, Decoding Budget=24K2026.05 | 64.73 | 75.69 | 65.44 | 56.79 | |
| CoRDSCase=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=3K, Decoding Budget=3K2026.05 | 64.31 | 74.73 | 64.46 | 57.51 | |
| Full KVCase=-, Streaming MC=-, Compression QA=-, Prefill Budget=50K, Decoding Budget=50K2026.05 | 63.93 | 74.68 | 62.11 | 55 | |
| InfiniPot-VCase=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=12K, Decoding Budget=12K2026.05 | 63.59 | 74.22 | 62.68 | 53.89 | |
| InfiniPot-VCase=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=24K, Decoding Budget=24K2026.05 | 63.52 | 74.22 | 63.22 | 53.11 | |
| SnapKV [24]Case=Case 1, Streaming MC=✗, Compression QA=✗, Prefill Budget=50K, Decoding Budget=3K2026.05 | 63.07 | 74 | 61 | 54.22 | |
| SnapKV [24]Case=Case 1, Streaming MC=✗, Compression QA=✗, Prefill Budget=50K, Decoding Budget=6K2026.05 | 63.03 | 74.22 | 60.55 | 54.33 | |
| SnapKV‡Case=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=24K, Decoding Budget=24K2026.05 | 63 | 74.22 | 61 | 53.78 | |
| Qwen2.5-VL-64Input Frames=64, Token Retain Ratio (r)=100% (Full Performance)2026.05 | 62.9 | 74.2 | 62.7 | 51.8 | |
| ST-SimDiff-64Input Frames=64, Token Retain Ratio (r)=50%2026.05 | 62.9 | 74.2 | 62.1 | 52.2 | |
| InfiniPot [21]Case=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=24K, Decoding Budget=24K2026.05 | 62.78 | 74.22 | 60.55 | 53.56 | |
| InfiniPot-VCase=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=6K, Decoding Budget=6K2026.05 | 62.78 | 74.11 | 60.78 | 53.44 | |
| ST-SimDiff-64Input Frames=64, Token Retain Ratio (r)=30%2026.05 | 62.4 | 73.8 | 61.7 | 51.8 | |
| UniformCase=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=24K, Decoding Budget=24K2026.05 | 62.22 | 74.22 | 59.22 | 53.22 | |
| SnapKV‡Case=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=12K, Decoding Budget=12K2026.05 | 62.07 | 74.44 | 58.89 | 52.89 | |
| InfiniPot [21]Case=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=12K, Decoding Budget=12K2026.05 | 61.67 | 74 | 57.78 | 53.22 | |
| InfiniPot-VCase=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=3K, Decoding Budget=3K2026.05 | 61.11 | 73.89 | 57.78 | 51.78 | |
| UniformCase=Case 2, Streaming MC=✗, Compression QA=✓, Prefill Budget=50K, Decoding Budget=6K2026.05 | 60.96 | 72 | 58.78 | 52.11 | |
| SnapKV†Case=Case 2, Streaming MC=✗, Compression QA=✓, Prefill Budget=50K, Decoding Budget=6K2026.05 | 60.63 | 72.11 | 57.56 | 52.22 | |
| InfiniPot [21]Case=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=6K, Decoding Budget=6K2026.05 | 60.52 | 72.89 | 57.33 | 51.33 | |
| UniformCase=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=12K, Decoding Budget=12K2026.05 | 60.26 | 74 | 55.33 | 51.44 | |
| Video-TwGresolution=High-Resolution (HR)2026.02 | 59.7 | 71.9 | 57.1 | 50 | |
| SnapKV‡Case=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=6K, Decoding Budget=6K2026.05 | 59.55 | 72 | 55.33 | 51.33 | |
| UniformCase=Case 2, Streaming MC=✗, Compression QA=✓, Prefill Budget=50K, Decoding Budget=3K2026.05 | 58.18 | 70.33 | 54.67 | 49.55 | |
| UniformCase=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=6K, Decoding Budget=6K2026.05 | 58.11 | 72.33 | 53.33 | 48.67 | |
| SnapKV†Case=Case 2, Streaming MC=✗, Compression QA=✓, Prefill Budget=50K, Decoding Budget=3K2026.05 | 57.89 | 69 | 54 | 50.67 | |
| InfiniPot [21]Case=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=3K, Decoding Budget=3K2026.05 | 57.55 | 67.11 | 54.55 | 51 | |
| Video-R1-7Breproduced=true2026.02 | 57.3 | 69.3 | 53.6 | 49.1 | |
| Qwen2.5-VL-7Bresolution=High-Resolution (HR)2026.02 | 57.2 | 71.2 | 53.8 | 46.7 | |
| SnapKV‡Case=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=3K, Decoding Budget=3K2026.05 | 56.26 | 66.67 | 52.22 | 49.89 | |
| Video-XL-7B2026.02 | 55.5 | 64 | 53.2 | 49.2 | |
| UniformCase=Case 3 (CKV), Streaming MC=✓, Compression QA=✓, Prefill Budget=3K, Decoding Budget=3K2026.05 | 55.48 | 66 | 52.44 | 48 | |
| FastV [4] (L=2)Case=Case 1, Streaming MC=✗, Compression QA=✗, Prefill Budget=48/6K (R=5.8)2026.05 | 55 | 59.67 | 54.55 | 50.78 | |
| VideoChat-R1-7Breproduced=true2026.02 | 54.4 | — | — | — | |
| Video-TwGresolution=Low-Resolution (LR)2026.02 | 53.6 | 63.7 | 49.4 | 47.7 | |
| Long-LLaVA-7B2026.02 | 52.9 | 61.9 | 51.4 | 45.4 | |
| LongVA-7B2026.02 | 52.6 | 61.1 | 50.4 | 46.2 | |
| FastV [4] (L=2)Case=Case 1, Streaming MC=✗, Compression QA=✗, Prefill Budget=48/3K (R=2.8)2026.05 | 50.96 | 54.11 | 50.11 | 48.67 | |
| LongLLaVA-9B2026.02 | 50.9 | 59.6 | 50.3 | 42.7 | |
| VideoTree2026.02 | 48 | 55.5 | 49.2 | 39.3 | |
| VideoAgent2026.02 | 46.9 | 53.3 | 49.7 | 37.8 | |
| Qwen2.5-VL-7Bresolution=Low-Resolution (LR)2026.02 | 46.6 | 56.2 | 43.2 | 40.4 | |
| Video-LLaVA-7B2026.02 | 39.9 | 45.3 | 38 | 36.2 | |
| ShareGPT4Video-8B2026.02 | 39.9 | 48.3 | 36.3 | 35 | |
| VideoChat2-7B2026.02 | 39.5 | — | — | 33.2 | |
| Long-VILA-R1-7Breproduced=true2026.02 | 34.6 | — | — | — |