Reflection: Definition and Examples

Reflection (or Reflexion) is an advanced paradigm in artificial intelligence that allows an agent or language model to self-evaluate after generating a response. Rather than producing a single, final output, the model examines its own production, identifies errors, inconsistencies, or gaps, and then generates an improved version. This process can be repeated over several iterations until a satisfactory quality level is reached.

This approach is directly inspired by human metacognitive mechanisms: when we solve a complex problem, we naturally check our reasoning, spot logical flaws, and adjust our response. Reflection transposes this behavior to AI systems by introducing an explicit feedback loop in the generation process.

In practice, reflection can be implemented in several ways. In prompt engineering, the model is explicitly asked to critique its own response before finalizing it. In more elaborate architectures like the Reflexion framework (Shinn et al., 2023), an agent receives feedback signals—from an environment, an external evaluator, or its own analysis—which it stores in memory to guide subsequent attempts.

Reflection is particularly effective for complex reasoning tasks, mathematical problem solving, code generation, and argumentative writing. It significantly reduces hallucinations and logical errors, making the model a more reliable and self-correcting system.