CoT

A prompting technique that forces LLMs to output intermediate reasoning steps to solve complex logic and math problems.

Jason Wei and the Google Brain team published the definitive CoT paper in 2022. The technique prompts LLMs to generate a sequence of intermediate logical steps before reaching a conclusion. It solves the performance gap in multi-step tasks: on the GSM8K math benchmark, PaLM 540B jumped from 18 percent to 57 percent accuracy using CoT. This method (often triggered by the phrase: Let's think step by step) makes the model's reasoning transparent and verifiable. It is a standard requirement for high-accuracy workflows in GPT-4 and Claude 3.5 Sonnet.

https://arxiv.org/abs/2201.11903