A paper posted on arXiv frames question answering over radiology reports as involving two separate decisions: whether the report contains an answer at all, and which span in the text provides that answer. The authors argue that these should not be collapsed into a single extraction step.

The motivation is to address two distinct failure modes. A false answer presents an unsupported claim as fact, while over-abstention hides what the report actually says. By dissociating answerability from extraction, the work aims to let a model abstain when no answer exists without suppressing valid answers when they do.

The abstract does not report experimental results, so the proposal is best read as a conceptual reframing of the task. Still, it highlights a practical tension for LLMs in clinical settings: knowing when not to answer may be as important as knowing where the answer lies.