Reflection / self-critique
- Reasoning
- Research term
Reflection, or self-critique, is asking an AI to review an answer it already gave and revise it if needed. The review can look for errors, gaps, contradictions, or a poor fit with your instructions.
The second pass runs on the same model, with the knowledge and assumptions behind the first answer, so an error rooted in them tends to survive. A review that brings in something new, such as a source document, an error message, or a failing test, gives the model a fixed point to compare against.
Where you’ll meet it
You're doing your taxes and ask your assistant to total a year of medical receipts. When you ask it to double-check, it returns the same total with a more careful-sounding explanation. One pharmacy payment is still counted twice.
What to do
Attach your bank or card statement, and ask it to match every receipt in its total to one line there and to list any receipt it couldn't match or matched twice. Look over that list yourself before you file.
Easy to mix up
Self-consistency makes several separate tries and puts them to a vote. Reflection makes one try, then reviews that try and revises it.
Check yourself
You ask your assistant for a cover letter and say to leave out salary. Its letter has two mistakes, and you reply only "Review your letter and fix any problems." Which mistake is the review more likely to fix?
Go deeperResearch found that unaided rechecks can turn right answers wrong
In Reflexion, an agent that fails a task writes notes on what went wrong, drawing on outside results such as test outcomes, and reads them on its next try. Self-Refine loops feedback and revision on one answer. Both report gains on some tasks. Huang and colleagues found in 2023 that without outside feedback, models rechecking their reasoning often did worse.
To decide whether a second pass belongs in a workflow, run the workflow on test cases with known answers, with and without it, and compare. Keep it for the kinds of error it catches, and count the cases it makes worse.