Research map / Multi-agent systems and games / LLM agents

ClinTutor-R1: Advancing Scalable and Robust One-to-Many Alignment in Clinical Socratic Education

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Builds ClinEdu, a multi-agent simulator for clinical ward-round Socratic tutoring, generates the ClinTeach dataset, and trains ClinTutor-R1 (Qwen2.5-VL-7B with SFT, LoRA, then GRPO with a rubric reward) for one-to-many clinical tutoring.

Problem

LLM tutors suffer context dilution and goal misalignment when one instructor must guide a group of heterogeneous trainees simultaneously, degrading sharply as group size grows; real clinical ward-round data is hard to obtain.

Approach

A persona-driven multi-agent simulator (ClinEdu) with decoupled patient scripts and personas, Specialist and Safety review agents, and a think-before-speaking tutor that writes per-student and group analyses, trained with SFT followed by GRPO using a three-axis rubric reward with a safety/format veto.

Key results

• ClinTutor-R1 reaches 8.35 average on MedXpertQA interactive evaluation versus 6.96 for base Qwen2.5VL (about 20% gain) and 8.42 for o3.
• In a 200-student real user study, ClinTutor-R1 averages 8.61 versus 8.42 for o3 and 5.67 for Qwen.
• Average Correction Success Rate of 88.50% in controlled error injection experiments.

Surprising findings

• Removing the thinking mechanism alone yields only a small drop in some metrics, while removing RL causes the largest drop, per ablation tables.
• Without the reward veto, the Medical Professionalism and Safety score drops sharply (6.92 on MedXpertQA), showing the veto's importance.

Limitations

• Simulator-based training may not capture emotional, noisy dynamics of real clinical wards; vision scope limited to text and static images.
• User study is short-term; longitudinal learning gains not measured.
• Explicit thinking increases latency that grows roughly linearly with student count.

Experimental setup

Datasets: MedXpertQA ClinTeach

Benchmarks: MedXpertQA MVME MMMU PMC-VQA ClinEdu interactive simulation

Metrics: ETS MSM MPS average score (1-10) IQ IE OR Correction Success Rate Cohen's kappa

Baselines: LLaVA-v1.6-Mistral-7B Qwen2.5VL-7B-Instruct InternVL3.5-8B-Instruct o3 GPT-4o DeepSeek-R1 EduChat-R1 TutorRL DRLTutor Med-SocraticLM

Builds on: GRPO SocraticLM TutorRL MedXpertQA

Closest papers by meaning

• MMedAgent-RL: Optimizing Multi-Agent Collaboration for Multimodal Medical Reasoning (ICLR 2026 · similarity 0.87)
• MASTER: Enhancing Large Language Model via Multi-Agent Simulated Teaching (NeurIPS 2025 · similarity 0.86)
• SocraticLM: Exploring Socratic Personalized Teaching with Large Language Models (NeurIPS 2024 · similarity 0.86)
• MACPO: Weak-to-Strong Alignment via Multi-Agent Contrastive Preference Optimization (ICLR 2025 · similarity 0.85)
• Learning “Partner-Aware” Collaborators in Multi-Party Collaboration (NeurIPS 2025 · similarity 0.84)
• CoRe: Collaborative Reasoning via Cross Teaching (ICML 2026 · similarity 0.84)