Title: ClinTutor-R1: Enhancing Scalable and Resilient One-to-Many Alignment in Clinical Socratic Education

Abstract:

Although Large Language Models (LLMs) have demonstrated exceptional proficiency in dyadic, one-on-one instruction, they encounter substantial obstacles in achieving effective one-to-many alignment. This challenge is particularly pronounced in settings like clinical ward rounds, where a single instructor must simultaneously direct a heterogeneous group of trainees. Existing models frequently struggle with context dilution and goal misalignment, unable to effectively reconcile the need for individualized scaffolding with the collective advancement of the group.

To overcome these limitations, we present ClinEdu, a multi-agent pedagogical simulator designed to capture the intricacies of group dynamics. Utilizing this simulation platform, we have compiled ClinTeach, a comprehensive dataset featuring Socratic teaching dialogues. Furthermore, we propose ClinTutor-R1, a novel vision-language agent specifically engineered for one-to-many alignment in clinical education. ClinTutor-R1 utilizes an explicit internal reasoning mechanism to track both individual belief states and group consensus.

We evaluated our framework through a rigorous protocol encompassing static benchmarks, in-situ interactive assessments within ClinEdu, expert reviews, and a user study involving 200 participants. Our findings indicate that ClinTutor-R1 surpasses base models by more than 20% and matches the performance of proprietary models. Additionally, the system demonstrates strong scalability, preserving high instructional quality even as student cohorts expand.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC