Improve an LLM judge
Preview
Define a rubric, add golden examples, and check the judge against human labels.
An LLM judge is a model call configured to score another model's work. It is useful when success requires judgment, such as whether a response follows a policy, explains a decision clearly or handles a customer appropriately. Improving its prompt does not require training the judge model's weights.
The preview supports rubric authoring, selecting several reviewed labels as development examples, independent validation fixtures, and exact published grader/candidate check comparison. Adoption is a separate reviewed action so a saved candidate cannot silently change learning settings.
Define what the judge should see and return
A judge needs:
- A rubric with concrete criteria and explanations of ambiguous cases.
- A scoring scale or labels with clear meanings.
- The task, candidate response and relevant conversation or tool evidence.
- Optional task-specific reference information, such as an expected answer.
- A selected judge model and, optionally, golden examples.
For example: “Pass only if the response checks eligibility before promising a refund. Offering to check eligibility is allowed; an unconditional promise is not.” Use the same criterion when humans review responses.
Golden examples and judge tests
| Example use | What the judge sees |
|---|---|
| Prompt demonstration | A sample response, its human label and an explanation of the judgment |
| Judge check | A response to score; its human label is withheld and used for comparison afterward |
Golden examples demonstrate how to apply the rubric. A judge's check dataset can contain many more examples than its prompt. Do not insert every labeled response into every call. Keep independent examples outside prompt refinement so you can check whether the judge applies the rules to unfamiliar cases.
Refine using reviewed examples
- Collect labels against the current rubric, including clear passes and failures.
- Run the judge on those responses and compare its scores with the human labels.
- Inspect disagreements. Resolve ambiguous labels before assuming the judge needs changing.
- Propose clearer instructions or selected golden examples for recurring mistakes.
- Test the candidate judge on previous cases and new independent cases. Check false passes and false failures for each criterion, not only overall agreement.
- Review the prompt/example changes and publish a new version if justified.
Adding a fixture tests the judge; it does not automatically add that example to the prompt. Adding a golden example to the prompt changes the judge and requires a new reviewed version.
Use it during training
Once selected for a run, the judge scores fresh attempts using that fixed version. Its scores supply the reward signal. New tasks do not rewrite the rubric, and an old response's human score cannot score a newly generated answer.
If your new labels reveal a missing requirement, improve the judge before training against it. Otherwise, the model can be rewarded for behavior you consider wrong. The judge's references and demonstrations are not automatically shown to the model performing the task.
See Change a grader safely before using an updated judge in continual learning.