Abstention

Abstention is an explicit decision not to provide a substantive prediction or answer for a case. It can reflect inadequate evidence, an unresolved conflict, a confidence threshold, or a policy restriction. An abstention should say which condition applied when that can be disclosed. It is distinct from a negative factual claim: 'not found in the reviewed sources' describes a bounded search result, while 'unknown' leaves the value unsettled.

Evaluation should report coverage, the fraction of cases answered, alongside error rates on answered cases. Increasing abstention can improve the apparent accuracy of the remaining answers without improving retrieval or evidence quality. Thresholds should reflect the cost of a wrong answer and the cost of leaving a case unresolved. Human review paths need to preserve the original evidence and the reason the system withheld an answer.

For an organization that must defend values, abstention prevents unsupported certainty from silently entering a record. It is useful only when unresolved cases remain visible and actionable. Calibration may inform confidence-based thresholds, but evidence gaps and conflicts often require rules beyond a numerical score. Null semantics describes the meaning of absent or negative values; abstention is the decision behavior applied when those states do not warrant an answer.

A refusal that says nothing is of little use. A good decline states why: no source in scope, sources in conflict, evidence too old, or the question needs a permission the user lacks. It may offer what is known, such as the conflicting values with their sources, and what would resolve the question. These forms keep the system honest about the difference between unknown and false, a distinction explained under null semantics and in the article on unknown versus false in open-world systems.

Abstention can fail in both directions. A system that never abstains will answer from weak evidence. A system that abstains too often is unusable, and users learn to bypass it. Set thresholds according to the cost of each error, and review a sample of abstentions to see whether an answer was available. For agents that update stored knowledge, a related decision is whether to update at all when new information conflicts with the stored value, covered in the article on agent memory updates.

Test with questions the corpus cannot answer, questions with conflicting sources, and questions with outdated evidence. Check that the system declines or qualifies and gives a reason. Measure the abstention rate and review a sample for unnecessary declines. Check that thresholds were set from calibrated confidence on labelled data.