Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

Large language models struggle with multi-step arithmetic, commonsense, and symbolic reasoning, and standard prompting gives the model no room to work through intermediate steps.

Not an architecture change: PROMPT with a few chain-of-thought EXEMPLARS in which the worked solution is written out as a series of intermediate reasoning steps, so the model imitates the step-by-step format at inference. Reasoning ability is elicited emergently by scale rather than trained.

NOT A SPEEDUP PAPER - and the cost implication is the opposite of one. Prompting a 540B-parameter model with EIGHT chain-of-thought exemplars achieves state-of-the-art accuracy on GSM8K, surpassing even FINETUNED GPT-3 WITH A VERIFIER. For this repo the operative consequence is the COST: every intermediate reasoning step is an OUTPUT token, and output tokens are the bandwidth-bound resource that dominates decode latency. CoT is therefore a latency and cost multiplier imposed by the prompt, not a free accuracy win, and this is the origin point for that accounting.

ADOPTION: in-production beyond dispute - chain-of-thought prompting is universal in deployed systems and needs no engine support at all. VENUE: NeurIPS 2022; the arXiv comment field does not state it, so venue rests on the widely-attested NeurIPS 2022 publication and should be re-verified against Semantic Scholar, which was rate-limited at verification time. INCLUDED AS THE COST-ORIGIN RECORD for the inference-time-compute sub-area: every downstream record (s1 budget forcing, self-consistency, best-of-N) multiplies the token count this paper introduced, and the reason reasoning tokens are expensive is precisely that they are output tokens on the memory-bandwidth-bound decode path. Deliberately hardware_relevance = [] because the technique is hardware-independent; the hardware dependence enters only through the token count it creates. AFFILIATIONS FILLED 2026-10-04 (paper-verification pass): from the Semantic Scholar graph API (paper/arXiv:2201.11903), which is the only machine-readable source that carries author affiliations — the arXiv Atom feed returns author names only, and no ORCID is exposed by either API. One known upstream typo corrected: "Tel Aviv Univesity" -> "Tel Aviv University". S2 affiliation coverage is partial (120 of 249 corpus records carry any), so these lists may be incomplete rather than exhaustive. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. This is STRUCTURAL rather than researched: neurips has no Findings volume and no workshop track in which a paper could appear, so 'main' is the only possible honest value and no per-paper evidence was needed. The field is annotated here so that 'findings' and 'workshop' remain MEANINGFUL - a filter for findings=true must return only the papers that actually are findings, which requires every other paper to be explicitly NOT findings.