At fixed mask ratios on generated completions from all four benchmarks, downstream-biased masking yields higher average target-token log-probability on masked positions than uniform random masking, whereas upstream-biased masking yields substantially lower confidence. The separation grows with stronger priority bias; the paper interprets downstream subproblems as better posed because upstream reasoning context remains visible.

Source: paper-fixed:PDF p. 2, Figure 1 and accompanying paragraphs; PDF p. 5, Figure 2 and Section 3.3