Feng He Zijun Chen Xinnian Liang Tingting Ma Yunqi Qiu Shuangzhi Wu Junchi Yan

Abstract

Recent advances in Large Reasoning Models (LRMs) trained with LongChain-of-Thought (Long CoT) reasoning have demonstrated remarkable cross-domaingeneralization capabilities. However, the underlying mechanisms supporting suchtransfer remain poorly understood. We hypothesize that cross-domaingeneralization arises from shared abstract reasoning prototypes -- fundamentalreasoning patterns that capture the essence of problems across domains. Theseprototypes minimize the nuances of the representation, revealing that seeminglydiverse tasks are grounded in shared reasoning structures.Based on thishypothesis, we propose ProtoReasoning, a framework that enhances the reasoningability of LLMs by leveraging scalable and verifiable prototypicalrepresentations (Prolog for logical reasoning, PDDL forplanning).ProtoReasoning features: (1) an automated prototype constructionpipeline that transforms problems into corresponding prototype representations;(2) a comprehensive verification system providing reliable feedback throughProlog/PDDL interpreters; (3) the scalability to synthesize problemsarbitrarily within prototype space while ensuring correctness. Extensiveexperiments show that ProtoReasoning achieves 4.7% improvement over baselinemodels on logical reasoning (Enigmata-Eval), 6.3% improvement on planningtasks, 4.0% improvement on general reasoning (MMLU) and 1.0% on mathematics(AIME24). Significantly, our ablation studies confirm that learning inprototype space also demonstrates enhanced generalization to structurallysimilar problems compared to training solely on natural languagerepresentations, validating our hypothesis that reasoning prototypes serve asthe foundation for generalizable reasoning in large language models.