Understanding the role of randomness when solving locally checkable labeling (LCL) problems in the LOCAL model has been one of the top priorities in the research on distributed graph algorithms in recent years. For LCL problems in bounded-degree graphs, it is known that randomness cannot help more than polynomially, except in one case: if the deterministic complexity of an LCL problem is in $Ω(\log n)$ and its randomized complexity is in $o(\log n)$, then the randomized complexity is guaranteed to be $poly(\log \log n)$. However, the fundamental question of \emph{which} problems with a deterministic complexity of $Ω(\log n)$ can be solved exponentially faster using randomization still remains wide open. We make a step towards answering this question by studying a simple, but natural class of LCL problems: so-called degree splitting problems. These problems come in two varieties: coloring problems where the edges of a graph have to be colored with $2$ colors and orientation problems where each edge needs to be oriented. For $Δ$-regular graphs (where $Δ=O(1)$), we obtain the following results. - We gave an exact characterization of the randomized complexity of all problems where the edges need to be colored with two colors, say red and blue, and which have a deterministic complexity of $O(\log n)$. - For edge orientation problems, we give a partial characterization of the problems that have a randomized complexity of $Ω(\log n)$ and the problems that have a randomized complexity of $poly\log\log n$. While our results are cleanest to state for the $Δ$-regular case, all our algorithms naturally generalize to nodes of any degree $d<Δ$ in general graphs of maximum degree $Δ$.