CK-20 Sampling-consistency hallucination detection (SelfCheckGPT; self-consistency check)
Draw several stochastic responses to the same prompt and flag statements that the other samples do not support, and produce sentence-level consistency scores.
What it does
Draws several stochastic responses to the same prompt and flags statements that the other samples do not support. Artefact: sentence-level consistency scores. Read the lowest-scoring sentences.
When to use it
Use when no source document is available to check a model answer against and you need a screen for likely invented statements.
When it does not fit
Needs no external database. It does not check against the source document, and consistent errors pass.
Result format
The result is a table with one row per sentence of the response, giving its consistency score across the sampled responses. It lists the lowest-scoring sentences as the statements the other samples do not support.
Requirements on the material
| Property | Requirement type | Requirement |
|---|---|---|
| Content type | Always applies | requires=text |
| Unit of analysis | Always applies | requires=claim |
| 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 | at least 5 stochastic samples per prompt |
|---|---|
| Additional sources needed | none |
| Not for (rules) | none |
| Not for (text) | Wrong choice as the only check, because it does not compare with the source document and consistent errors pass. It screens for invention and does not show truth. |
Choosing and ordering
| Kind of technique | Verification technique, runs in step 6 |
|---|---|
| Phase tag | validate |
| Other phase tags | none |
| Kind of finding | contradiction |
| Question answered | what |
| Strongest statement it supports | describe |
| Needs as input |
|
| Makes as output |
|
| Must exist first | None |
| Loop target | proceed with limit stated |
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: repeatable with fixed settings.
| Stakes | Check | Person involved | Counts as a different method |
|---|---|---|---|
| exploratory | spot check | optional | no |
| exploratory | repeat runs | none | no |
| decision-informing | spot check | optional | no |
| decision-informing | repeat runs | none | no |
| decision-informing | human review sample | needed; left out below high stakes | no |
| decision-informing | source-span citation | none | no |
| high-stakes | spot check | optional | no |
| high-stakes | repeat runs | none | no |
| high-stakes | human review sample | needed | no |
| high-stakes | source-span citation | none | no |
| high-stakes | second technique with different assumptions | none | conditional |
| high-stakes | independent review | needed | no |
Failure modes and limits
| Model suitability | model plus code |
|---|---|
| Typical model failure | Repeats the same error consistently, so the samples agree. |
| Error sources | input data; language model |
| Certainty limits | Certainty drops because the samples come from the same model and share its errors (risk of bias) and because tests covered one text type only (indirectness). |
Related techniques
- should follow PE-12 Model summarisation of documents (abstractive summarisation; multi-document summarisation): generated summaries are the usual subject
- should follow PE-13 Retrieval-augmented analysis (RAG (retrieval-augmented generation); grounded question answering over a corpus): generated answers are the usual subject
- pairs with CK-17 Source-span and quotation check by code (quote-and-verify; extractive grounding; citation verification): a source-span check finds consistent errors that this check passes
Evidence and tag basis
Tested. Tested on GPT-3 biographies with manual annotation (2023, abstract). (Manakul, Liusie and Gales 2023, arXiv 2303.08896)
Tag basis: input profile inference; selection inference; sequencing inference; robustness inference. All tag values are inference unless a source is named.