> This effect cannot be explained by memorization since < 1.41% of the initial puzzle board states appear in our training set.