However, they hardly compare against the conventional next-token-prediction objective, which suggests they don't beat it, except for highly non-linear tasks like sudoku. Unfortunately, machine learning papers often tend to focus exclusively on the positives without mentioning bad aspects.