Examples

1. 1.

   Wei et al. (2022) at Google Brain showed that adding 'Let's think step by step' to math word problems improved PaLM's GSM8K accuracy from 17.9% to 58.1% without any fine-tuning.

2. 2.

   DeepSeek R1 uses chain-of-thought reasoning natively - producing long, explicit reasoning traces before answers, a technique adopted in OpenAI's o1/o3 series for difficult maths and coding tasks.

3. 3.

   A legal AI contract-review tool prompts Claude 3.5 with a structured CoT sequence: 'First identify the governing law clause, then assess force majeure scope, then evaluate limitation-of-liability caps, then summarise risk.'