Selective Prediction on To See or To Please Diagnostic
73.5Baseline AccuracyQwen2.5-VL-72B
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Qwen2.5-VL-72BStrategy=Diagnostic-Guided Selective Prediction, Coverage=50%2026.03 | 73.5 | 77.8 | 4.3 | |
| Qwen2.5-VL-7BStrategy=Diagnostic-Guided Selective Prediction, Coverage=50%2026.03 | 72.1 | 81.6 | 9.5 | |
| Molmo2-4BStrategy=Diagnostic-Guided Selective Prediction, Coverage=50%2026.03 | 71.3 | 76.6 | 5.3 | |
| Llama-3.2-11BStrategy=Diagnostic-Guided Selective Prediction, Coverage=50%2026.03 | 69.9 | 69.9 | 0 | |
| Phi-3.5-VisionStrategy=Diagnostic-Guided Selective Prediction, Coverage=50%2026.03 | 69 | 69.7 | 0.7 | |
| LLaVA-NeXT-7BStrategy=Diagnostic-Guided Selective Prediction, Coverage=50%2026.03 | 68.1 | 75.4 | 7.3 | |
| Pixtral-12BStrategy=Diagnostic-Guided Selective Prediction, Coverage=50%2026.03 | 66.9 | 66.1 | 0.8 |