List your company

Hallucination Detection on PM-seen vs. hallucination (N=300)

99.4AUROCOracle margin
90.97693.16395.3597.537May 7, 2026

Evaluation results

MethodLinks
Oracle margin2026.0599.4100
Entropy (output-based)Source=output-based2026.0596.70.5
Linear probe (post-hoc)Evaluation Protocol=post-hoc2026.0596.4—
MLP head (post-hoc)Evaluation Protocol=post-hoc2026.0595.318
MLP head (end-to-end)Evaluation Protocol=end-to-end2026.0591.30.5