Prefill-stage hallucination risk detection on Benchmark-500 Relaxed Consensus (Pvote ≥ 0.8)
0.6957AUROC (Mean)Risk-Cos
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Risk-Cosgeneration_mode=prefill only2026.03 | 0.6957 | 0.61 | |
| Risk-Margingeneration_mode=prefill only2026.03 | 0.673 | 0.59 | |
| Risk-Entropygeneration_mode=prefill only2026.03 | 0.502 | 0.42 | |
| Risk-Lossgeneration_mode=prefill only2026.03 | 0.326 | 0.25 |