REIN Framework for Reasoning and Reliability Alignment in LRMs

REIN mitigates reasoning and knowledge hallucinations in Large Reasoning Models through a structured reflection-and-abstention alignment. The framework forces models to follow a think-reflect-answer sequence and utilizes a reward mechanism to encourage explicit abstention when no valid reasoning chain is found.