Minimum Sufficient Reasoning: How Much Chain-of-Thought Does Training Data Need?

Reasoning models are typically trained on long chain-of-thought traces, even for problems that need only a few steps. A natural alternative is to keep the shortest correct trace per problem. However, this has been studied almost entirely on math, and a short trace can reach the correct answer by luck. We introduce Minimum Sufficient Reasoning (MSR). A Qwen3.6-35B teacher generates traces at five depths, from direct answers to full extended thinking, on GSM8K, MATH, ARC-Challenge, and StrategyQA. We then keep the shallowest depth whose k-sample pass rate clears a per format reliability threshold. Under a pure-guessing null, this statistical selector caps the chance that a guesser survives selection at about 2% and requires no learned judge. We find that (1) required depth varies sharply with domain and difficulty; (2) contamination in naive shortest-correct selection tracks answer-space size, with 33.2% of selections on yes/no StrategyQA and 12.1% on four-choice ARC statistically indistinguishable from guessing, compared to near zero on free-form math; and (3) MSR reduces training-data tokens by 96.0% relative to uniformly long traces. Without any fine-tuning, Qwen3.5-4B prompted with a few MSR examples produces much shorter outputs at similar accuracy.