The idea in one paragraph.
Prompting a large model with a few worked examples that show intermediate steps elicits step-by-step reasoning and sharply improves multi-step tasks.
THE MACHINERY
Chain-of-thought exemplars
Demonstrations include the reasoning steps, not just the answer.
Spreading compute over tokens
Buy more serial computation by emitting more intermediate tokens.
What the paper reports.
Findings are reproduced as concise corpus summaries, with their original confidence labels. They are not independent replications.
Prompting PaLM 540B with eight chain-of-thought exemplars reaches state-of-the-art accuracy on GSM8K, surpassing finetuned GPT-3 with a verifier.
Inspect a source excerpt
prompting a PaLM 540B with just eight chain-of-thought exemplars achieves state-of-the-art accuracy on the GSM8K benchmark of math word problems,…Read the passage in context on arXiv ↗
The gains are emergent with scale and do not appear in small models.
LIMITATIONS / CORPUS READING
Where to keep your judgement.
- The reasoning trace is not verified; a fluent chain can still reach a wrong answer.
- No feedback path - the model cannot check its own steps against anything external.
Why it belongs in this atlas.
Expands the output space rather than the input space. The talk frames this as smearing the computation over many more tokens instead of demanding the answer in one.