Grounded Multi-Video Question Answering on WikiVideo
94Ref PrecisionCRAFT
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| CRAFTBase VLM=Qwen3-VL-30B-Instruct, Atomic Claims=true, Number of Frames=1282026.05 | 94 | 62 | 73.5 | 85.5 | 85.8 | 84.8 | 80.9 | |
| CRAFTBase VLM=Qwen3-VL-30B-Instruct, Frame Sampling=DKS, Number of Frames=322026.05 | 94 | 83.2 | 79.7 | 96.6 | 64.7 | 76.1 | 82.4 | |
| CRAFTBase VLM=Qwen3-VL-30B-Instruct, Frame Sampling=uniform, Number of Frames=322026.05 | 93 | 64 | 74.6 | 84.5 | 84.4 | 83 | 80.5 | |
| CRAFTBase VLM=Qwen3-VL-30B-Instruct, ASR access=true, Number of Frames=1282026.05 | 87.1 | 84.9 | 85.4 | 94.9 | 65.6 | 76.2 | 82.3 | |
| CRAFTBase VLM=Qwen3-VL-30B-Instruct, Critic Loop=true, Number of Frames=1282026.05 | 85.9 | 84.5 | 84.2 | 95.3 | 66.8 | 77.3 | 82.2 | |
| CRAFTBase VLM=Qwen3-VL-30B-Instruct, Configuration=Baseline, Number of Frames=1282026.05 | 83.3 | 83.4 | 83.4 | 95.1 | 66.2 | 76.4 | 81.4 | |
| InternVL-3.5-30B-A3BNumber of Frames=128, ASR access=true2026.05 | 81.5 | 84.8 | 83.1 | 74.3 | 71.2 | 72.7 | 77.9 | |
| InternVL-3.5-30B-A3BNumber of Frames=128, ASR access=false2026.05 | 80.2 | 82.1 | 81.1 | 73.1 | 68.9 | 71 | 76.1 | |
| Gemma-4-31BNumber of Frames=128, ASR access=true2026.05 | 73.2 | 77.8 | 75.4 | 62.9 | 65.1 | 64 | 69.7 | |
| Gemma-4-31BNumber of Frames=128, ASR access=false2026.05 | 72.1 | 74.8 | 73.4 | 61.8 | 63 | 62.4 | 67.9 | |
| Molmo2-8BNumber of Frames=1282026.05 | 64.1 | 68.2 | 66.1 | 51.2 | 59.8 | 55.2 | 60.7 |