Answering the Unanswerable Is to Err Knowingly: Analyzing and Mitigating Abstention Failures in Large Reasoning Models

AI systems can recognize problems but won't admit it: The research reveals that large reasoning models actually possess the internal capability to identify when questions are unanswerable (such as math problems missing critical information), but they fail to communicate this limitation to users. This represents a critical gap between what the AI "knows" internally and what it tells users.

Current AI models frequently provide false answers instead of saying "I don't know": When faced with unanswerable questions, these reasoning models continue to generate responses rather than appropriately abstaining or declining to answer. This behavior undermines user trust and could lead to harmful decisions based on fabricated reasoning.

A practical solution exists without major system overhauls: The researchers developed a lightweight two-stage method combining "cognitive monitoring" with intervention techniques that significantly improves when AI systems appropriately decline to answer, while maintaining their ability to solve solvable problems. This suggests responsible AI improvements don't necessarily require complete system redesigns.

Trustworthiness requires honest uncertainty communication: The paper frames appropriate abstention as essential for "trustworthy AI," emphasizing that users need to know the limits of AI reasoning. An AI system that confidently provides wrong answers is less trustworthy than one that acknowledges when it lacks sufficient information.

This is a solvable alignment problem: The research demonstrates that improving AI honesty about its limitations is technically feasible, suggesting that responsible AI development can address this specific failure mode through targeted interventions at inference time.