<ul><li>VeriThinker is introduced as a novel approach for compressing Chain-of-Thought (CoT) reasoning, which helps in making Large Reasoning Models (LRMs) more efficient.</li><li>Unlike traditional methods, VeriThinker fine-tunes LRMs solely through an auxiliary verification task instead of using synthetic concise CoT data, leading to a reduction in overthinking and unnecessary reasoning chain lengths.</li><li>Experiments show that VeriThinker significantly reduces reasoning chain lengths while maintaining or slightly improving accuracy when applied to tasks like MATH500 and AIME25, showcasing its effectiveness in enhancing model efficiency.</li><li>VeriThinker can also be generalized to speculative reasoning and the code is available at https://github.com/czg1225/VeriThinker.</li></ul>

VeriThinker: Learning to Verify Makes Reasoning Model Efficient

Discover more