Abstention
A decision or operational practice of withholding a prediction or answer—such as responding “I don't know” or declining to make a judgment—when uncertainty or decision risk is high.
Example
When evidence is insufficient, a system states that it cannot determine the answer rather than guessing.
Distinctions and cautions
Abstention primarily addresses uncertainty and the risk of error. Its motivation differs from a Refusal based on safety rules or usage policy. Uncertainty Estimation and Calibration can inform a decision to abstain, but neither is abstention itself.
Open questions and cautions
The threshold of uncertainty or risk at which withholding an answer is appropriate varies by use case and by the cost of an incorrect answer.
Evidence
View evidence →
Selective Classification for Deep Neural Networks
- Title
- Selective Classification for Deep Neural Networks
- Author / organization
- Geifman & El-Yaniv
- Year
- 2017
- Source type / publication status
- NeurIPS peer-reviewed research; published
- Supported concepts / claim context
- Abstention/reject option; risk–coverage trade-off.
- Scope and limitations
- Classification setting; application to free-form LLM abstention is a conceptual extension.
- Access / version note
- Published conference proceedings.
- Recorded official URL / DOI / identifier
- Official URL / recorded link
Evaluating large language models for accuracy incentivizes hallucinations
- Title
- Evaluating large language models for accuracy incentivizes hallucinations
- Author / organization
- Kalai et al.
- Year
- 2026
- Source type / publication status
- Nature peer-reviewed research; published
- Supported concepts / claim context
- Accuracy, Abstention, Hallucination; evaluation incentives.
- Scope and limitations
- Formal model and selected frontier-model case study. Utility depends on the use case.
- Access / version note
- Published Nature article.
- Recorded official URL / DOI / identifier
- DOI:
10.1038/s41586-026-10549-w(DOI resolver)