Large language models often hallucinate when forced to answer questions they do not know. A new arXiv preprint proposes a simple fix: let the model abstain. The authors describe abstention as a way to reduce hallucination by skipping questions the model would likely get wrong.
Crucially, this abstention is implemented outside the model itself, using a technique called selective answering. A confidence threshold determines when the model should decline to respond. This makes the behavior controllable—operators can tune how cautious the model is without retraining or modifying its weights.
The approach is notable because it does not require learning to abstain. Instead of teaching the model a new skill, the method leverages existing confidence signals. The abstract does not provide experimental details, but the framing suggests a practical, lightweight alternative to more invasive hallucination-reduction strategies.