Review loop
- Agents
- Standard term
A review loop is a cycle in which an AI result is checked against stated criteria, each failure is sent back with its evidence for revision, and a set limit decides when the cycle stops.
Evidence tied to a criterion tells the reviser what failed and gives the next check a stable target. Returning a whole draft with “make it better” leaves the intended change less clear. A later pass still needs checking because fixing one part can affect another.
Where you’ll meet it
An assistant has tailored your résumé to a job posting that lists six requirements. The draft reads well, and when you hold it against the posting, two of the requirements appear nowhere in it.
What to do
- When addressing each missing requirement, quote the posting and include only experience and qualifications you can support.
- Check the new draft against all six again, since a fix can drop something else.
- After three rounds, keep the best accurate draft and resolve remaining gaps yourself; leave unsupported qualifications out.
Easy to mix up
In reflection, the model reconsiders its own work, often with no new evidence. A review loop checks the work against stated criteria and sends each failure back with its evidence.
Check yourself
An AI's product description must stay under 40 words and name all six features from the spec. Round after round, every draft under 40 words leaves features out, and every draft with all six runs long. What does that pattern suggest?
Go deeperA model rechecking its own reasoning often misses its mistakes
In a 2023 study, Huang and colleagues found that models asked to recheck their own reasoning, with no outside feedback, often failed to improve it, and sometimes changed a right answer to a wrong one.
A script can check a word limit or required fields. A source comparison can establish whether a claim matches that source, but the source itself may need verification. Tone can be judged against examples and criteria; no single checker settles every property.
Sources