Chain-of-Thought
Make the model show its work and watch hard problems start to crack.
Grounded in the research on chain-of-thought prompting
what it is
Chain-of-thought is a prompting trick: instead of demanding the final answer, you make the model write out its reasoning step by step before it commits. On a math word problem, a normal prompt blurts a number and often whiffs. A chain-of-thought prompt walks through it line by line. The reasoning isn't just for show. Spelling out intermediate steps is what lets the model land the right answer.
who named it
Jason Wei and a Google Brain team dropped the paper in January 2022, 'Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,' published at NeurIPS that November. Their move: feed the model just eight worked examples that included the reasoning, not only answers. On GSM8K, a grade-school math benchmark, the 540-billion-parameter PaLM model jumped to state of the art, beating a fine-tuned GPT-3 with a verifier.