Uncertainty Estimation
A family of methods for estimating or quantifying how uncertain a model's output is, using information such as probability distributions, differences across repeated outputs, semantic variation, or ensembles combining multiple outputs or methods.
Aliases: Uncertainty Quantification; UQ.
Example
Repeated answers to the same question sometimes converge on nearly the same response and sometimes diverge substantially. That difference is used as a signal of uncertainty.
Distinctions and cautions
Confidence is a score or expression attached to a particular prediction; uncertainty estimation is a method or process for obtaining information about uncertainty. Whether estimated uncertainty corresponds appropriately to the actual likelihood of error is a calibration question.
Open questions and cautions
Different methods measure different types of uncertainty. It has not been established to what extent their values can be compared directly across free-form responses, long-form text, multilingual settings, and agentic use.
Evidence
View evidence →
Bayesian Prompt Ensembles: Model Uncertainty Estimation for Black-Box Large Language Models
- Title
- Bayesian Prompt Ensembles: Model Uncertainty Estimation for Black-Box Large Language Models
- Author / organization
- Tonolini et al.
- Year
- 2024
- Source type / publication status
- Findings ACL peer-reviewed research; published
- Supported concepts / claim context
- Uncertainty Estimation; prompt uncertainty.
- Scope and limitations
- Mainly natural-language classification; uses a small labeled validation set.
- Access / version note
- Published ACL Findings.
- Recorded official URL / DOI / identifier
- DOI:
10.18653/v1/2024.findings-acl.728(DOI resolver)
Detecting hallucinations in large language models using semantic entropy
- Title
- Detecting hallucinations in large language models using semantic entropy
- Author / organization
- Farquhar et al.
- Year
- 2024
- Source type / publication status
- Nature peer-reviewed research; published
- Supported concepts / claim context
- Uncertainty Estimation, Abstention, Hallucination.
- Scope and limitations
- Requires sampling and clustering by semantic equivalence; cost and applicability depend on the setting.
- Access / version note
- Published Nature article.
- Recorded official URL / DOI / identifier
- DOI:
10.1038/s41586-024-07421-0(DOI resolver)