CReaTe-LM: A large language model for stepwise, heuristic clinical reasoning instruction

Effective clinical reasoning is central to competent medical practice, yet cultivating this skill remains one of the most challenging goals in medical education. Although large language models (LLMs) create new opportunities, most follow a passive answering paradigm - where users pose queries and models passively respond - failing to support the step-by-step, heuristic instruction essential for clinical reasoning. To bridge this gap, we propose CReaTe-LM, the first language model that effectively provides heuristic and stepwise clinical reasoning instruction. We begin by introducing a unified Clinical Reasoning Instructional Scaffold (CRIS) and designing a multi-agent cooperation pipeline involving three specialized LLM agents - Teacher, Expert, and Student - to emulate authentic instructional dynamics. On this basis, we construct CRID, the first large-scale, high-quality Chinese dataset for clinical reasoning instruction, consisting of over 6.4 K multi-turn dialogues and 31 million tokens. CReaTe-LM is then trained via supervised fine-tuning and further enhanced through reinforcement learning targeting three key instructional abilities: structured instruction, heuristic guidance, and personalized feedback. For evaluation, we propose a comprehensive, multi-perspective framework combining LLM-based evaluation, expert human assessment, and standard Natural Language Processing (NLP) metrics. Extensive experiments show that CReaTe-LM consistently outperforms eight state-of-the-art baselines. In particular, on a 0-100 scale, it surpasses GPT-4.1 by 12.7, 15.0, and 10.4 points in LLM-based evaluation, human assessment, and NLP metrics, respectively. In an authentic classroom study, CReaTe-LM delivers reliable gains over a standardized, instructor-led lecture, supporting its practical instructional value. These results support CReaTe-LM's effectiveness for structured, adaptive, and heuristic clinical instruction.