Chain-of-Thought Prompting Elicits Reasoning in…

This paper asks why showing a language model a few worked examples of step-by-step reasoning unlocks problems that direct prompting cannot touch. Read it to see how eight hand-written exemplars lifted PaLM 540B from 17.9% to 56.9% on grade-school math, why meaningless filler tokens explain none of the gain, and why the same prompt actively hurts smaller models.

Reading focus: How a few worked examples make a large model write its own intermediate steps before committing to an answer. Why the ablations show the gain comes from natural-language steps before the answer, not extra tokens or activated knowledge. Why the ability emerges only near hundred-billion-parameter scale, and what fluent but illogical chains do to smaller models.

NeurIPS 2022. Wei, Wang, Schuurmans, Bosma, Ichter, Xia, Chi, Le, and Zhou. 30 min read, easy difficulty.