Generated knowledge prompting is a two-stage technique in which a language model first generates relevant background statements about a question and then answers the question conditioned on those statements. In prompting, it belongs to the family of methods that make a model externalize intermediate content before committing to a final answer, but unlike chain-of-thought it produces facts rather than reasoning steps.
The method was introduced in 2022 by researchers at the University of Washington and the Allen Institute for AI for commonsense reasoning benchmarks. Its central insight is that a model often holds useful knowledge it does not apply spontaneously, and that eliciting that knowledge explicitly can improve accuracy without retrieval infrastructure or fine-tuning.
How generated knowledge prompting works
The technique separates knowledge generation from knowledge integration. In the first stage, a few-shot prompt shows several example questions, each paired with a short factual statement, and then asks the model to produce similar statements for the new question. In the second stage, each generated statement is placed before the question, and a model answers it.
In the original formulation, the system samples many knowledge statements per question, often around 20, answers the question once with each statement and once without any, and selects the answer that receives the highest model confidence across all variants. This selection rule means that a single highly relevant statement can decide the outcome, while irrelevant statements are simply outvoted.
The table summarizes the stages and the main design choices at each one.
| Stage | Input | Output | Key design choice |
|---|---|---|---|
| Knowledge generation | Question plus few-shot demonstrations | Several short factual statements | Number of statements and sampling temperature |
| Knowledge integration | One statement prepended to the question | Candidate answer with confidence | Same or different answering model |
| Answer selection | All candidate answers | Final answer | Highest confidence or majority vote |
Design parameters that shape results
Results depend heavily on the quality and format of the demonstrations used to elicit knowledge. Demonstrations should show concise, declarative statements that are relevant to the question type but do not state the answer directly, because statements that leak answers teach the model to guess rather than to inform.
| Parameter | Typical setting | Effect when increased |
|---|---|---|
| Number of statements | 5 to 20 | Better coverage, higher token cost |
| Sampling temperature | Moderate, about 0.5 to 0.8 | More diverse but less reliable facts |
| Statement length | One or two sentences | More detail, more room for error |
| Demonstrations | 3 to 5 examples per task | Better format control, longer prompt |
With current chat models, practitioners often use a single-call variant: the prompt asks the model to list relevant facts and then answer in the same response. This is cheaper and simpler, but it loses the independent sampling and voting that made the original method robust.
Comparison with related prompting techniques
Generated knowledge prompting is closest to step-back prompting and to retrieval-augmented generation, but the source and purpose of the added context differ. Step-back prompting asks for a general principle, retrieval brings in external documents, and generated knowledge draws specific facts from the model’s own parameters.
| Technique | What the model adds | Source of new information |
|---|---|---|
| Generated knowledge prompting | Relevant factual statements | Model parameters only |
| Chain-of-thought prompting | Step-by-step reasoning | Model parameters only |
| Step-back prompting | A higher-level principle or concept | Model parameters only |
| Self-ask prompting | Follow-up subquestions and answers | Model parameters or a search tool |
| Retrieval-augmented generation | Retrieved passages | External documents |
These techniques combine well. A common pattern is to generate knowledge first and then request chain-of-thought reasoning over it, or to merge generated statements with retrieved passages so that the model fills gaps that the document index does not cover. Self-consistency can also be applied by voting across answers produced from different statements.
When to use generated knowledge prompting
The technique is most useful for questions that require implicit world knowledge that a model possesses but tends to skip, such as commonsense reasoning, numerical common sense and everyday science. It helps less when the task depends on private, recent or highly specialized information that the model never learned.
| Situation | Recommended approach | Reason |
|---|---|---|
| Commonsense or general science questions | Generated knowledge prompting | Knowledge exists in the model but is not applied |
| Company policies or private data | Retrieval-augmented generation | The model cannot know the facts |
| Multi-step arithmetic or logic | Chain-of-thought prompting | The bottleneck is reasoning, not facts |
| Fast-changing news or prices | Retrieval or tool use | Parametric knowledge is outdated |
Limitations and failure modes
The main risk is that generated knowledge can be wrong. A model that hallucinates a plausible but false statement may then answer confidently on that basis, so the technique can amplify errors as well as correct them. Sampling several statements and voting reduces, but does not remove, this risk.
- Higher cost and latency: the full method requires many model calls per question.
- Answer leakage: statements that simply restate a guessed answer add no knowledge and bias the result.
- Diminishing returns: very large models already apply much of their knowledge, so gains tend to be smaller than on earlier models.
- No verification: unlike retrieval, the added facts have no external source that can be checked or cited.
- Prompt sensitivity: small changes to the demonstrations can change which facts the model produces.
For high-stakes use, generated statements should be treated as hypotheses to verify, for example by pairing the method with chain-of-verification or by checking key claims against a trusted source.
The idea in brief
Before answering a tricky question, a careful person often says what they already know about the subject. Generated knowledge prompting asks an AI model to do the same. First the model writes a few short facts about the topic. Then it uses those facts to answer.
The key takeaway is that the model often knows more than it uses. Asking it to recall facts first helps it apply that knowledge. The facts still come from the model itself, so they can be wrong and deserve a check.
A practical scenario
A team building a tutoring assistant for middle school science notices that the model sometimes gives wrong answers to commonsense questions such as whether a metal spoon or a wooden spoon gets hot faster in boiling soup. The team adds a knowledge generation step with four demonstrations that each pair a question with one or two neutral facts, such as how well common materials conduct heat.
For each student question, the assistant samples eight short statements, answers once with each statement and once without, and keeps the answer with the highest average confidence. On an internal test set of 300 questions, accuracy rises by several percentage points, and the generated facts are shown to students as a brief explanation.
The outcome is more accurate and better explained answers, at the cost of roughly nine model calls per question, which the team accepts because responses are cached for repeated questions.