A preprint finds mixed SFT scored higher than next-chunk reasoning after RLVR, which used over 60 times more GPU hours in the study.