Tag
This paper investigates why large language models answer structurally unanswerable questions, concluding that models recognize impossibility but fail to route that recognition to refusal, highlighting a routing failure rather than an encoding failure in their decision-making.