Mode-hopping persists under the continuous soft answer-probability difference P(correct) − P(incorrect), across the five non-generative prompting evaluation families and both model families. Thus the observed phenomenon is not confined to discontinuities in hard accuracy. · CiteArk
Not assessedPlan blockedFindingsoft-metric-mode-hopping
Mode-hopping persists under the continuous soft answer-probability difference P(correct) − P(incorrect), across the five non-generative prompting evaluation families and both model families. Thus the observed phenomenon is not confined to discontinuities in hard accuracy.