REIN: Bridging the Gap Between Reasoning and Reliability via Reflection and Abstention Alignment

REIN is an alignment framework that reduces hallucination in large reasoning models through reflection and abstention alignment. It trains models to explicitly self-reflect before the final answer and to encourage abstention (e.g., "I don't know") when none of the sampled reasoning chains contains the correct answer. Across multiple backbone models, hallucination proxy metrics decrease by 58–72% relative to the base models, average coverage remains at 86–91%, and selective accuracy on attempted questions improves by 6.6–14.2%.