Selective Prediction on ScienceQA 60 (test)
0.99AUROCRidge Arclength
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Ridge ArclengthModel=SmolVLM-Instruct, Variant=PR-dim/Ridge Proximity2026.06 | 0.99 | 0.146 | 0.85 | 0.119 | |
| Ridge ArclengthModel=Idefics3-8B-Llama3, Variant=Spherized/Ridge Proximity2026.06 | 0.934 | 0.14 | 0.856 | 0.11 | |
| Baseline kNNModel=Idefics3-8B-Llama3, Variant=kNN-R² proximity2026.06 | 0.907 | 0.163 | 0.833 | 0.117 | |
| logP(x)Model=SmolVLM-Instruct, Variant=Sequence logP(x)2026.06 | 0.79 | 0.243 | 0.753 | 0.169 | |
| Flow MatchingModel=SmolVLM-Instruct, Variant=Neg dist. to correct centroid2026.06 | 0.769 | 0.326 | 0.67 | 0.175 | |
| Topology LID-MLEModel=Idefics3-8B-Llama3, Variant=Neg LID-MLE2026.06 | 0.511 | 0.526 | 0.479 | 0.214 |