Long-form Video Question Answering on LongVideoBench (test)
57.3Long Video AccuracyQwen3-VL-4B
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Qwen3-VL-4BPipeline=Monolithic Baselines, K (Frame Pool Size)=144, Model Size=4B2026.05 | 57.3 | 64.1 | 1,493 | |
| GP-4BPipeline=Single-Model GridProbe, Selector Model=Qwen3-VL-4B, QA Model=Qwen3-VL-4B, K (Frame Pool Size)=1442026.05 | 55.9 | 62.4 | 575 | |
| Qwen3-VL-8BPipeline=Monolithic Baselines, K (Frame Pool Size)=144, Model Size=8B2026.05 | 55.1 | 62.7 | 2,574 | |
| GP-2B->4BPipeline=Cross-Model GridProbe, Selector Model=Qwen3-VL-2B, QA Model=Qwen3-VL-4B, K (Frame Pool Size)=1442026.05 | 54.3 | 60.4 | 452 | |
| GP-8BPipeline=Single-Model GridProbe, Selector Model=Qwen3-VL-8B, QA Model=Qwen3-VL-8B, K (Frame Pool Size)=1442026.05 | 53.4 | 60.7 | 1,068 | |
| GP-2B->8BPipeline=Cross-Model GridProbe, Selector Model=Qwen3-VL-2B, QA Model=Qwen3-VL-8B, K (Frame Pool Size)=1442026.05 | 52 | 59.7 | 735 | |
| GP-2BPipeline=Single-Model GridProbe, Selector Model=Qwen3-VL-2B, QA Model=Qwen3-VL-2B, K (Frame Pool Size)=1442026.05 | 51.4 | 57.3 | 301 | |
| Qwen3-VL-2BPipeline=Monolithic Baselines, K (Frame Pool Size)=144, Model Size=2B2026.05 | 49.8 | 56.4 | 868 | |
| Uniform-Meff -> 8BPipeline=Cross-Model GridProbe (Control), Selection=Uniform, QA Model=Qwen3-VL-8B, K (Frame Pool Size)=1442026.05 | 49.8 | 58.5 | 735 |