[Interesting techniques]

• Sample cot traces using greedy decoding.
• The model arrived at the right answer even without right thinking steps coincidentally.

[Future research]

• The paper shows that COT shows negligible improvements in small language models.