List your company

Question Answering on TruthfulQA (AUROC, Correct mean, Incorrect mean)

0.858AUROCOracle margin δ
0.647920.702460.7570.81154May 7, 2026

Evaluation results

MethodLinks
Oracle margin δ2026.050.85862.7118.3
Full hidden-state probe2026.050.689——
Entropy H2026.050.6561.442.02