Reflection with LLM: How to Make AI Review Its Own Work

Reflection refers to a structured process where the model evaluates and improves its own output. It typically involves generating an initial response, analyzing that response against defined goals or constraints, and producing a refined version based on detected issues if any. This approach enhances accuracy, consistency, and reliability by enabling self-assessment before finalizing results.

Read this Story for free: https://medium.com/towards-artificial-intelligence/reflection-with-llm-how-to-make-ai-review-its-own-work-2db122fca1d8?sk=8ed3a8d93a61be41ed428343e2f34952

Let’s understand this with an example.
Consider a scenario where LLM is given a dataset containing sales transactions and is asked to answer a business question, such as “Show total sales by region for Widget A in February 2024.”

In this example, the model first generates Python (pandas) code to query the dataset, then evaluates the correctness of its own output, refines the code if necessary, and finally validates the result using a feedback mechanism. The following code demonstrates this single-pass reflection process — from code generation to refinement and validation.

Data Preparation and Schema Definition