CK-22 LLM-as-judge with human calibration (model-graded evaluation)
Use a strong model to grade outputs against a rubric and check the judge against human grades on a subsample, and produce judge scores with judge-human agreement.
What it does
Uses a strong model to grade outputs against a rubric, then checks the judge against human grades on a subsample. Artefact: judge scores with judge-human agreement. Read the agreement before using any score.
When to use it
Use when many model outputs must be graded against a rubric and you need to know if the grading model agrees with people.
When it does not fit
The result concerns chatbot answer preference, not extraction accuracy or numerical verification, so check the judge on the task at hand.
Result format
The result gives the rubric, the judge's scores on the outputs and the judge-human agreement measured on a human-graded subsample. It states the agreement first, before the scores, and says whether the judge can be used for the task.
Requirements on the material
| Property | Requirement type | Requirement |
|---|---|---|
| Content type | Always applies | requires=text |
| Unit of analysis | Always applies | requires=document |
| Appraises source credibility | Applies in some cases | no |
Not restricted on: Structure, Variable types, Time structure, How the sample was chosen, Level of detail available, Number of sources, Whether sources can be joined, Amount of material, Known data problems, Producer interest, Source dependence, Origin, Document genre, Language, Personal information, Security classification, Licence.
| Minimum material | a written rubric and human grades on a subsample, usually 100 or more items |
|---|---|
| Additional sources needed | reference structure |
| Not for (rules) | none |
| Not for (text) | Wrong choice for extraction accuracy or numerical verification, because the evidence concerns chatbot answer preference. The judge must be checked on the task at hand. |
Choosing and ordering
| Kind of technique | Verification technique, runs in step 6 |
|---|---|
| Phase tag | validate |
| Other phase tags | none |
| Kind of finding | option ranking |
| Question answered | which option |
| Strongest statement it supports | describe |
| Needs as input |
|
| Makes as output |
|
| Must exist first | None |
| Loop target | return to analyse |
Library checks by stakes
Each stakes level includes the checks of the level below it. Checks that need a person apply at high stakes only, so the process leaves them out at lower stakes (decision 38). Repeatability: judgement dependent.
| Stakes | Check | Person involved | Counts as a different method |
|---|---|---|---|
| exploratory | spot check | optional | no |
| decision-informing | spot check | optional | no |
| decision-informing | gold standard agreement | conditional; left out below high stakes | no |
| decision-informing | prompt variation | none | no |
| high-stakes | spot check | optional | no |
| high-stakes | gold standard agreement | conditional | no |
| high-stakes | prompt variation | none | no |
| high-stakes | human review sample | needed | no |
| high-stakes | independent review | needed | no |
Failure modes and limits
| Model suitability | model plus code |
|---|---|
| Typical model failure | Favours longer, first-listed or own-family answers. |
| Error sources | input data; language model; coder or rater judgement |
| Certainty limits | Certainty drops when the judge shows position, length or self-preference bias (risk of bias) and when the task differs from the one the judge was tested on (indirectness). |
Related techniques
- should follow PE-12 Model summarisation of documents (abstractive summarisation; multi-document summarisation): summaries are typical outputs graded
- should follow PE-13 Retrieval-augmented analysis (RAG (retrieval-augmented generation); grounded question answering over a corpus): retrieval answers are typical outputs graded
- pairs with CK-04 Gold-standard validation sample (ground truth set; human-coded validation set): human grades come from a human-coded set
Evidence and tag basis
Tested. Strong judges agreed with human preferences over 80 per cent of the time, the same level as between humans, and showed position, verbosity and self-enhancement bias (2023, abstract). (Zheng et al. 2023, arXiv 2306.05685)
Tag basis: input profile inference; selection inference; sequencing inference; robustness inference. All tag values are inference unless a source is named.