REIN Framework Boosts LLM Reasoning Reliability and Reduces Hallucinations

Large Reasoning Models (LRMs) often struggle with hallucinations, producing incorrect or unsupported information, which limits their safe deployment. A new framework called REIN (Reasoning and Reliability via Reflection and Abstention Alignment) aims to enhance LRM reliability by tackling two distinct types of hallucinations: reasoning flaws and factual knowledge gaps. To combat reasoning hallucinations, REIN trains models to generate a structured sequence of thought, allowing for explicit self-reflection before committing to a final answer. For knowledge hallucinations, where the model lacks necessary factual information, REIN incorporates a reward system that encourages the model to explicitly abstain (e.g., by saying "I don't know") rather than fabricating an answer. Extensive evaluations on mathematical and commonsense reasoning benchmarks show that REIN consistently improves selective accuracy and significantly reduces instances where models confidently provide incorrect answers. It achieves these gains efficiently within a single forward pass, without needing complex external controls or multi-round critiques, making it a practical solution for more trustworthy AI.