Pretraining does not monotonically improve distinguishing truth from plausibility: performance on surprising truths and common misconceptions mode-hops in both OLMo3 and Apertus.

来源：paper:§3.3, Figure 5, p.4; Appendix G, Figure 21, p.17