# How do you ask an AI to critique and improve its own answer?

Date: 2024-04-10 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/ask-ai-critique-improve-own-answer/

**Answer:** Ask for a draft, then ask the model to grade that draft against a short list of criteria you define, quoting the weak parts, and finally to rewrite it fixing only what the critique found. Stop after one or two rounds, or as soon as every criterion passes.

A first answer from a chatbot is a first draft, and like any first draft it usually has fixable problems: a vague opening, a missing point, a paragraph that runs too long. You can point these out yourself, round by round. But you can also ask the model to do the first pass of review, as long as you tell it exactly what to review against.

This technique has a research name. In a 2023 paper called Self-Refine, researchers had one language model write an output, give itself feedback, and then revise using that feedback, all with the same model and no extra training. Across seven tasks, from rewriting reviews to math reasoning, people and automatic measures preferred the refined outputs over first drafts from GPT-3.5 and GPT-4, by about 20 percentage points on average. The key word is feedback: the critique has to be specific enough to act on.

## Why "make it better" isn't enough

If you simply type "improve this," the model has no target. It will often make the text longer, more formal and more decorated, which may not be what you need. A rubric fixes that. A rubric is a short list of criteria, each with a clear test, that describes what a good answer looks like for this task. Writing it takes a minute, and it is also useful for you: it forces you to decide what you actually want.

## A worked loop in three prompts

Suppose a small bakery owner needs a short announcement that the shop will close for two weeks of renovations.

**Step 1: draft.**

```prompt
Write a 100-word announcement for our bakery's Facebook page. We close from [DATE] to [DATE] for renovations and reopen with a bigger seating area. Friendly tone.
```

**Step 2: critique against a rubric.**

```prompt
Before rewriting anything, review your draft against this rubric. For each criterion, give a score from 1 to 5, quote the exact words that cost points, and say in one sentence what would fix it. Do not rewrite yet.

1. The closing dates appear in the first two sentences.
2. Readers learn why we are closing and what is new when we reopen.
3. There is one clear action (for example, stock up before we close).
4. 100 words or fewer.
5. Warm and plain: no clichés like "exciting journey."
```

An illustrative critique might give criterion 1 a score of 2 ("the dates only appear in the last line") and criterion 5 a score of 3 (quoting "we're thrilled to embark on this exciting chapter").

**Step 3: revise only what failed.**

```prompt
Now rewrite the announcement. Fix only the criteria that scored below 4, and keep everything that scored 4 or 5. Then list each change you made in one line.
```

The list of changes at the end lets you check quickly that the model did what the critique said, and didn't quietly rewrite good parts.

## The stopping rule

Decide in advance when the loop ends. A practical rule: stop when every criterion scores 4 or higher, or after two critique rounds, whichever comes first. After that, the changes tend to be cosmetic, and each round costs you time to read. If a criterion still fails after two rounds, the problem usually isn't wording. The model is missing information, like the reopening date or the actual name of the new menu item, and only you can supply it.

> **Try it:** Take an email you need to send this week. Ask for a draft, write a four-point rubric for it, and run one critique and one revision. Compare the first and final versions side by side and note which rubric point made the biggest difference.

## When does self-critique make answers worse?

It works best on qualities the model can see in the text itself: length, structure, tone, whether a required point is present. It works much less well on whether a fact or a calculation is correct. A paper posted in October 2023, "Large Language Models Cannot Self-Correct Reasoning Yet," tested exactly this. When models were asked to review and fix their own reasoning without any outside information, they generally failed to improve, and sometimes changed correct answers into wrong ones. If a model didn't know a fact the first time, asking it to check itself won't supply the missing knowledge.

So for facts, figures, quotes and code, add outside feedback instead of relying on self-review. Check claims against a reliable source, run the code, or redo the math with a calculator, then paste what you found into the chat: "The test failed with this error" or "The official date is May 3." That is real feedback, and the model can use it well.

Watch, too, for over-polishing. Each round can strip out a little personality or add a little length. If the third version reads blander than the second, keep the second. The rubric is there to protect what you want, so if you notice the answer drifting, add the missing quality as a new criterion rather than running more rounds.

## Sources

- Self-Refine: Iterative Refinement with Self-Feedback: https://arxiv.org/abs/2303.17651v1
- Large Language Models Cannot Self-Correct Reasoning Yet: https://arxiv.org/abs/2310.01798v1