List your company

Prefill-stage hallucination risk detection on Benchmark-500 Strict Consensus Pvote = 1.0 vs. Clean

0.6939AUROC (Mean)Risk-Cos
0.3370760.4297130.522350.614987Mar 20, 2026

Evaluation results

MethodLinks
Risk-Cosgeneration_mode=prefill only2026.030.69390.58
Risk-Margingeneration_mode=prefill only2026.030.6850.57
Risk-Entropygeneration_mode=prefill only2026.030.53230.44
Risk-Lossgeneration_mode=prefill only2026.030.35080.25