Must the reviewer be a different model?

No. The same model can perform a separate review task, although that does not make its judgment independent or infallible.

Does reflection update the model's weights?

Not necessarily. Common workflows revise outputs or retain feedback in context without performing parameter training.

How do we know reflection helped?

Compare independently assessed results before and after the review process, including both repairs and newly introduced errors, on representative examples.

AI Glossary / Generative AI and language / Intermediate

Reflection

Definition

Reflection is a workflow strategy in which an AI system examines an earlier output or action result before deciding what to do next. The examination may identify omissions, inconsistencies, failed tests, or departures from a stated goal, and may lead to revision. In this context, reflection describes an engineered review process. It does not imply consciousness, human self-awareness, or an ability to reliably detect every mistake.

What it means

A first draft and a review have different purposes. The draft attempts the task; the review compares the attempt with criteria. An AI workflow can separate these stages even when the same model performs both.

A reflection step may receive the original request, candidate output, a checklist, and evidence such as test results or source material. Useful feedback identifies a specific problem and a possible correction instead of merely declaring that the answer is good.

Reflection is broader than any one research framework. For example, Self-Refine studies iterative feedback and revision, while Reflexion uses feedback and reflective text across agent trials. Their names should not be treated as interchangeable implementation specifications.

How it works

First define what the output must accomplish and what evidence can establish success. Generate an initial attempt, then ask a reviewer component to inspect it against those criteria. The reviewer may be the same model, another model, a tool, a human, or a combination.

Use the resulting feedback to revise, accept, retry, or escalate. Set a stopping rule, such as a maximum number of revisions or a requirement to pass specified checks. Repeated self-approval is not a substitute for external evidence.

Keep the original and revised outputs when evaluating the workflow. A revision can introduce new errors, so assess the final result and compare against a baseline without reflection. Additional calls consume time and resources even when they do not improve quality.

Three practical examples

1. A report-writing workflow reviews a draft against a required-section checklist before returning it to an editor.

2. A coding assistant examines a failed test result, proposes a targeted correction, and reruns the relevant checks before accepting the change.

3. A document-answering system compares a draft's claims with retrieved source passages and flags statements for which it cannot identify support.

Think of it this way

Think of reflection as a separate editing pass after writing a draft. The change of task encourages attention to weaknesses that the drafting pass may have missed.

The analogy has limits. A model does not become an independent expert merely because it receives an editing instruction. It may repeat its original assumptions, accept persuasive but false wording, or invent a problem. The quality of the criteria and evidence matters more than the label placed on the review.

Worked example

Consider a hypothetical evaluation of 100 draft answers. Independent reviewers find that 70 meet all requirements before reflection. After one review-and-revision cycle, 12 previously failing answers become acceptable, but 5 previously acceptable answers become unacceptable.

The final acceptable count is 70 + 12 − 5 = 77, giving a 7-percentage-point improvement. Reporting only the 12 repairs would hide the regressions. If the extra review costs two seconds per answer, it adds 200 seconds of processing time across the set, excluding other overhead.

These invented numbers illustrate how to assess the workflow. The team should also inspect error severity: replacing a minor style defect with a serious factual error would matter more than a simple pass-count comparison suggests.

Illustrative case study

Fictional case: Juniper Guides uses an assistant to draft equipment instructions. It adds a reflection step that checks whether every procedure includes prerequisites and a clear stopping condition.

The review catches some missing sections, but an editor notices that a revision also adds a plausible procedure absent from the approved manual. The team changes the review criteria to require source support for procedural additions and routes unsupported suggestions to a human.

It evaluates the revised workflow on held-out examples before adoption. The lesson is that reflection can improve a process only when its feedback and resulting changes are themselves assessed against the task's actual requirements.

Common misconceptions and mistakes

Reflection is not guaranteed self-correction. The same model may preserve the same misunderstanding in both draft and critique, and a second model can also be wrong.

It is not necessarily training. Many reflection workflows operate through prompts, feedback, and context without changing model weights. If feedback is retained across trials, the storage and retrieval mechanism should be stated explicitly.

Longer critiques are not automatically better critiques. A review should identify actionable issues linked to evidence. Finally, repeating revision indefinitely can waste resources or degrade an already adequate result, so stopping conditions are part of the design.

Business use

Reflection can be useful when outputs have explicit quality criteria and the review can access meaningful evidence. It may help catch omissions, formatting failures, or inconsistencies before work reaches a user.

Business assessment should compare final quality, regressions, latency, and cost with a simpler workflow. Decide which issues can be checked automatically and which need human judgment. Preserve concise records of detected problems and actual corrections so the team can determine whether the reflection stage contributes measurable value rather than merely producing reassuring commentary.

Check your understanding

A reviewer model says “all checks passed,” but no source or test supports that statement. Has the workflow demonstrated correctness?

Answer: No. The statement is another model output. The workflow needs evidence appropriate to its criteria and should verify any claimed checks that affect acceptance.

Related terms

Feedback
Self-Refine
Reflexion
agent workflow
critique
revision
evaluator
test-time computation
human review
stopping criterion

Learn next

Choose one narrow drafting task and define three observable acceptance criteria. Compare a single-pass baseline with one reflection cycle on the same held-out inputs. Have an independent reviewer record repairs, regressions, and unresolved errors. Retain reflection only if its benefit justifies the added resources and complexity.

Sources

Sources: Google Machine Learning Glossary, https://developers.google.com/machine-learning/glossary — primary definition of reflection in agentic workflows. Madaan and colleagues, Self-Refine, https://arxiv.org/abs/2303.17651 — original abstract on iterative self-feedback and revision. Shinn and colleagues, Reflexion, https://arxiv.org/abs/2303.11366 — original abstract distinguishing linguistic feedback and episodic memory from weight updates. These abstracts and the glossary section were read; illustrative results are invented teaching examples.

Previous: Referring expression generation Next: Reflection removal