ReasoningFreeA large model gets about 57% of grade-school math word problems right with greedy chain-of-thought decoding; you can't train it further but can spend much more compute at inference.
At higher temperature, the same problem yields several reasoning paths that don't always end on the same number.