Generated Knowledge: Definition and Examples

Generated Knowledge is an advanced prompt engineering technique introduced by Liu et al. in 2022. It is based on a simple yet powerful principle: before asking a language model to answer a question or perform a task, first ask it to generate relevant factual knowledge on the topic. This knowledge then serves as enriched context to produce a more accurate and informed response.

The process generally takes place in two distinct steps. In the first step (generation), the model is asked to produce facts, definitions, or explanations related to the subject. In the second step (integration), this generated knowledge is injected into a new prompt containing the final question. This approach allows the model to rely on explicit reasoning rather than drawing solely from its internal parameters.

This technique is particularly effective for common sense reasoning tasks, complex factual questions, and problems requiring deep contextualization. It differs from Retrieval-Augmented Generation (RAG) in that it does not use an external source: the model generates the knowledge itself from what it learned during training.

Generated Knowledge can be combined with other techniques such as Chain-of-Thought or Self-Consistency to achieve even more robust results. Its main advantage is that it requires no additional infrastructure—no vector database or retrieval pipeline—while significantly improving the reliability of responses on factual topics.