可监控性作为免费馈赠：RLVR如何自发对齐推理

Large Language Models (LLMs) excel at complex reasoning tasks, yet their internal mechanisms often remain opaque, hindering understanding and trust in their decision-making. This paper introduces a novel concept of "monitorability," aiming to reveal LLM reasoning processes through self-generated, interpretable intermediate steps. We propose the "Retrieval-based Verification and Refinement" (RLVR) framework, which spontaneously generates these monitorable steps without additional supervision by iteratively retrieving relevant information and refining its reasoning path. Experimental results demonstrate that RLVR not only significantly enhances LLM performance across tasks like multi-hop question answering, mathematical reasoning, and fact verification, but also produces highly interpretable and verifiable intermediate steps. These steps allow human users to easily trace and evaluate the model's reasoning process, thereby increasing trust in its outputs. RLVR's unique characteristic makes it a "free gift," as it naturally provides insights into the model's internal reasoning while simultaneously boosting performance, offering a new avenue for building more transparent and reliable AI systems. This work underscores the importance of designing self-explanatory AI systems and lays a foundation for the future development of trustworthy AI.