Paper, researchers, and primary source
Major lab research / reasoning
Chain-of-thought prompting shows that a few worked natural-language rationales can elicit multi-step reasoning from sufficiently large language models.
Contribution 1
Formalized few-shot chain-of-thought prompting.
Contribution 2
Demonstrated large gains on arithmetic, commonsense, and symbolic reasoning tasks.
Contribution 3
Observed that the technique becomes effective mainly at larger model scales.
Research context
reasoning / 2022
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models places chain of thought inside the broader reasoning discussion at Google Research, with prompting supplying a second analytical lens. Its contribution chain has three links: Formalized few-shot chain-of-thought prompting; Demonstrated large gains on arithmetic, commonsense, and symbolic reasoning tasks; and Observed that the technique becomes effective mainly at larger model scales. This framing makes reasoning a property to inspect within the study, not a label that settles later deployments. Its builder-facing implication is that worked examples can change reasoning behavior dramatically, but generated rationales are not guaranteed to be faithful explanations.
Methods and evidence reading
1 cataloged method notes
Method 1
The experimental design in Chain-of-Thought Prompting Elicits Reasoning in Large Language Models tests formalized few-shot chain-of-thought prompting and demonstrated large gains on arithmetic, commonsense, and symbolic reasoning tasks against the paper's documented baselines, datasets, model variants, or systems workloads.
How to read the evidence
A careful reading of Chain-of-Thought Prompting Elicits Reasoning in Large Language Models starts with the experiment's declared scope, not the reputation of Google Research. The editorial method record pairs two moves: Formalized few-shot chain-of-thought prompting; and Demonstrated large gains on arithmetic, commonsense, and symbolic reasoning tasks. The outcome-facing contribution is: Observed that the technique becomes effective mainly at larger model scales. This supports the bounded implication that worked examples can change reasoning behavior dramatically, but generated rationales are not guaranteed to be faithful explanations. It does not remove the source limit that transfer from Chain-of-Thought Prompting Elicits Reasoning in Large Language Models must retain or retest contamination controls, sampling policy, answer extraction, benchmark tasks, verifier behavior, and prompt design, because its chain of thought finding is bounded by the reported study. Follow-on evaluation should therefore vary prompting while retaining an explicit chain of thought baseline. An independent check of Chain-of-Thought Prompting Elicits Reasoning in Large Language Models needs a fixed chain of thought comparison, a declared prompting variation, and saved cases where Demonstrated large gains on arithmetic, commonsense, and symbolic reasoning tasks does not carry over.
Findings in the source record
1 paper-specific findings
- The reported evidence in Chain-of-Thought Prompting Elicits Reasoning in Large Language Models supports observed that the technique becomes effective mainly at larger model scales.
Practical implication for AI builders
Google Research / 2022
Proposed BrokenGPT application
Research blueprint / proposed status
Proposed route placement / chain of thought
For a proposed BrokenGPT experiment based on Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, maintain task-specific worked-example prompt packs and evaluate both final answers and tool-verifiable intermediate claims without presenting hidden reasoning as guaranteed evidence. Keep the chain of thought path isolated, versioned, and attributable to this research record.
Proposed acceptance test / prompting
Validate the proposed chain of thought route against the paper's reported outcome: Observed that the technique becomes effective mainly at larger model scales. The acceptance record for Chain-of-Thought Prompting Elicits Reasoning in Large Language Models should pair stability across samples, verifier agreement, and answer correctness with separate prompting failures, preventing one chain of thought average from settling the decision.
Proposed decision boundary / reasoning
Balance inference compute, faithfulness, and unresolved errors before promoting the proposed reasoning design. Because even if the reported result reproduces, domain shifts, open-ended conversations, language changes, unseen problem forms, and unfaithful rationales can reverse its product value and must be measured separately, adoption remains conditional on replay under BrokenGPT's selected model, runtime, and policy configuration.
Limitations, verification, and source
Boundaries recorded with the paper
Limitations
- Transfer from Chain-of-Thought Prompting Elicits Reasoning in Large Language Models must retain or retest contamination controls, sampling policy, answer extraction, benchmark tasks, verifier behavior, and prompt design, because its chain of thought finding is bounded by the reported study.
- Even if the reported result reproduces, domain shifts, open-ended conversations, language changes, unseen problem forms, and unfaithful rationales can reverse its product value and must be measured separately.
PRIMARY SOURCES
- 01Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
Google Research — Primary primary arXiv paper / 28 January 2022 / Jason Wei, Xuezhi Wang, Dale Schuurmans, and 6 more