Chain-of-Thought Prompting Elicits Reasoning in Large Language Models | Paper Notes

https://arxiv.org/abs/2201.11903

Introduced chain-of-thought (CoT) prompting.

Shows that chain-of-thought prompting leads to substantial performance improvements across various tasks, including arithmetic, commonsense, and symbolic reasoning. It enabled a PaLM 540B model to achieve state-of-the-art accuracy on the GSM8K benchmark of math word problems.

This "prompting only" approach is valuable because it doesn't require extensive training datasets or fine-tuning a separate model for each new task.

CoT prompt structure:

• Input: The problem or question presented to the language model.
• Chain of Thought: A sequence of natural language steps that show how to arrive at the solution. These steps explain the reasoning process, providing an "interpretable window" into its thought process.
• Output: The final answer to the problem.

The ablation study demonstrated that the specific sequential, natural language reasoning steps provided by CoT prompting are the reason for its effectiveness.