Google

Once a Large Language Model has generated feedback on its own output, that feedback can be used to construct a new prompt. This subsequent prompt instructs the LLM to revise and improve its original response, forming a key step in an iterative self-refinement cycle.

Using Generated Feedback to Prompt for Response Refinement

Imagine you are using a language model to get a summary of a complex topic. You have already received an initial response and then prompted the model to provide feedback on its own work.

**Initial Response:** 'Quantum entanglement is a physical phenomenon where the quantum state of each particle in a group cannot be described independently of the others. This non-local correlation means a measurement on one particle instantaneously influences the others, even at great distances, challenging classical intuitions.'

**Model's Self-Generated Feedback:** 'The summary is accurate but uses technical jargon like "non-local correlation" and "classical intuitions" that may not be clear to a general audience. The explanation could be simplified using an analogy.'

Based on the information above, write the *next* prompt you would use to instruct the model to improve its initial response.

Crafting a Refinement Prompt

A prompt template for refining a Large Language Model's output explicitly states the inputs the model will receive to guide the revision. For example, the model can be instructed with the following template: 'You will be provided with a question, its corresponding response, and some feedback on this response.' This structure ensures the LLM uses all three components—the original question, the initial response, and the feedback—to generate an improved answer.

Example of a Prompt Template for Response Refinement

Arrange the following actions into the correct logical sequence to guide a language model through one cycle of improving its own output.

An AI model provided an initial response to a prompt and was then instructed to generate feedback on its own work. Based on the information below, which follow-up prompt is best designed to guide the model toward a more comprehensive and refined answer?

**Initial Prompt:** "Summarize the main causes of the Roman Empire's decline."

**Initial Response:** "The Roman Empire fell mainly due to barbarian invasions."

**Generated Feedback:** "This response is overly simplistic. It correctly identifie

Learn Before

Related