A new preprint on arXiv tackles a persistent problem in retrieval systems: a high relevance score does not guarantee that an answer is correct. The authors propose a training-free abstention method that lets a system refrain from answering when confidence is low, based on the structure of the evidence graph rather than on additional training.
The core idea is to apply a boundary cut to the selected evidence within the evidence graph. This structural operation appears to identify when the supporting evidence is too weak or inconsistent to justify a confident answer, allowing the system to abstain instead of returning a potentially misleading result.
Because the method is training-free, it could be applied to existing retrieval pipelines without retraining or fine-tuning. The abstract is brief, so details on the exact boundary-cut procedure and evaluation are not yet available, but the approach points toward label-free abstention as a practical safeguard for structured retrieval.