Hallucination detection on SQuAD v2 (test held-out)
0.8577AUCHierarchicalTransformer
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| HierarchicalTransformerProbe Variant=M3, Status=Best Probe2026.04 | 0.8577 | 70.57 | 66.44 | |
| Token F1 Similaritytype=Token-overlap2026.04 | 0.4222 | 36.52 | 31.16 | |
| Cosine Similarity (Max)type=Embedding-based2026.04 | 0.411 | 42.69 | 51.46 | |
| Softmax Similaritytype=Embedding-based2026.04 | 0.4105 | 42.54 | 50.68 | |
| Mean Similaritytype=Embedding-based2026.04 | 0.4092 | 42.2 | 49 |