Source-linked AI summary
Chiron-o1: Igniting Multimodal Large Language Models towards Generalizable Medical Reasoning via Mentor-Intern Collaborative Search
Haoran Sun, Yankai Jiang, Wenjie Lou, Yujie Zhang, Wenjie Li, Lilong Wang, Mianxin Liu, Lei Liu, Xiaosong Wang
TL;DR
Medical MLLMs need better multimodal clinical reasoning data, but existing approaches lack a comprehensive way to search and evaluate effective reasoning paths. The paper proposes MICS, constructs MMRP, and trains Chiron-o1 with curriculum learning; Chiron-o1 achieves state-of-the-art performance across medical visual question answering and reasoning benchmarks. The approach remains costly and MMRP requires further expansion.
Problem
Existing approaches lack a comprehensive framework for searching and evaluating effective reasoning paths for medical CoT data.
Method
MICS uses mentor models to initialize reasoning, intern models to validate paths, and an MICS-Score to evaluate CoT quality before constructing MMRP and training Chiron-o1 with stage-wise SFT.
Results
Chiron-o1 outperforms state-of-the-art medical reasoning models across seven in-domain and out-of-domain benchmarks.
Takeaways & Limitations
MICS provides a collaborative strategy for constructing multimodal medical reasoning data, supporting Chiron-o1's robust reasoning performance across benchmarks.
Takeaways & Limitations
MICS requires numerous costly API requests and computational resources, while MMRP requires further expansion in scale.
Abstract
from arXiv · showhide
Multimodal large language models (MLLMs) have begun to demonstrate robust reasoning capabilities on general tasks, yet their application in the medical domain remains in its early stages. Constructing chain-of-thought (CoT) training data is essential for bolstering the reasoning abilities of medical MLLMs. However, existing approaches exhibit a deficiency in offering a comprehensive framework for searching and evaluating effective reasoning paths towards critical diagnosis. To address this challenge, we propose Mentor-Intern Collaborative Search (MICS), a novel reasoning-path searching scheme to generate rigorous and effective medical CoT data. MICS first leverages mentor models to initialize the reasoning, one step at a time, then prompts each intern model to continue the thinking along those initiated paths, and finally selects the optimal reasoning path according to the overall reasoning performance of multiple intern models. The reasoning performance is determined by an MICS-Score, which assesses the quality of generated reasoning paths. Eventually, we construct MMRP, a multi-task medical reasoning dataset with ranked difficulty, and Chiron-o1, a new medical MLLM devised via a curriculum learning strategy, with robust visual question-answering and generalizable reasoning capabilities. Extensive experiments demonstrate that Chiron-o1, trained on our CoT dataset constructed using MICS, achieves state-of-the-art performance across a list of medical visual question answering and reasoning benchmarks. Codes are available at https://github.com/manglu097/Chiron-o1
1 Introduction
Medical MLLMs often use direct prediction or RL-based optimization without fully addressing multimodal clinical reasoning and effective CoT construction. The paper introduces MICS and MMRP, then trains Chiron-o1 to achieve competitive reasoning performance across medical benchmarks.
- Motivation: Existing medical MLLMs often produce immediate answers while overlooking interleaved image-text reasoning needed for clinical scenarios.RL-based medical models also remain limited on real-world clinical questions and do not generate novel reasoning paradigms.
- Contributions: MICS searches for effective step-by-step CoT data by combining mentor-generated reasoning paths with intern-model validation.The strategy retains effective reasoning steps, discards low-quality or hallucinated ones, and uses an MICS-Score to assess CoT quality.
- Results: Seven-benchmark evaluation reports that Chiron-o1 outperforms state-of-the-art medical reasoning models across all evaluated benchmarks.The evaluation includes both in-domain and out-of-domain scenarios.
- Contributions: MMRP combines QA pairs, image-text alignment data, and reasoning paths across 12 imaging modalities and 20 body systems.
- Contributions: Chiron-o1 is a multimodal medical model designed to reason about complex clinical problems in both in-domain and out-of-domain settings.
2 Related Works
Prior work has improved MLLM reasoning through prompt tuning, supervised fine-tuning, reinforcement learning, and medical reasoning models, while multimodal medical reasoning remains challenging. CoT dataset construction methods guide reasoning through prompt-based, plan-based, or learning-based strategies.
- Reasoning in Medical MLLMs: Prompt tuning, supervised fine-tuning, and reinforcement learning are established techniques for improving MLLM reasoning capabilities.
- Reasoning in Medical MLLMs: Medical reasoning models such as HuatuoGPT-o1 and Baichuan-M1 target rigorous, transparent reasoning for medical problems.Integrating multimodal medical information into such reasoning remains an open challenge.
- Construction of CoT Datasets: CoT datasets for supervised fine-tuning emphasize reasoning processes rather than only final answers in multimodal reasoning.
- Construction of CoT Datasets: CoT construction research is organized into prompt-based, plan-based, and learning-based methods.
3 Methods
The method constructs multimodal medical reasoning data with MICS, then trains Chiron-o1 through curriculum-based supervised fine-tuning. MICS searches mentor-generated reasoning paths, evaluates them with intern models, and retains high-value paths for training.
- Mentor-Intern Collaborative Search: MICS collaboratively searches effective reasoning paths by combining multiple mentor models with intern-model feedback.Mentors generate intermediate steps, while smaller intern models validate whether the resulting paths support correct answers.
- Mentor-Intern Collaborative Search: Intern models evaluate each reasoning-path prefix by completing the reasoning process and measuring the frequency of correct answers.The proportion of interns answering correctly is used as the path’s MICS-Score.
- Mentor-Intern Collaborative Search: At each search iteration, MICS selects the highest-scoring reasoning path as the prefix for subsequent search until maximum depth or an early-stopping condition.Mentor competitiveness is calculated by multiplying value scores from previous search iterations.
- MMRP Construction: The resulting MMRP reasoning subset contains question, reasoning-path, and answer triplets for training multimodal medical reasoning.MMRP also includes QA and image-text alignment data spanning medical tasks and modalities.
- Chiron-o1 Training: Chiron-o1 is trained with curriculum-based SFT that progresses from text medical QA to increasingly complex multimodal medical tasks.The training objective combines predefined MMRP subsets in specified proportions.
4 Experiments and Results
Chiron-o1 is evaluated on medical VQA and reasoning benchmarks, with ablations testing MICS, training-set composition, and stage-wise curriculum training. It achieves strong benchmark performance, while the ablations show that MICS reasoning data, VQA data, and the complete training strategy each contribute under different task demands.
- Medical Reasoning Benchmarks: Chiron-o1-8B surpasses MedVLM-R1 by 27.2% in accuracy and 6.9% in semantic similarity on multimodal reasoning benchmarks.On pure-text reasoning, it exceeds MedReason and HuatuoGPT-o1 by 12.9% and 7%, respectively.
- Effect of MICS Strategy: Excluding reasoning data lowers average performance by 3%, while replacing MICS paths with vanilla mentor-generated paths lowers it by 2.7% across four benchmarks.The vanilla method bypasses intern-model evaluation during reasoning-path construction.
- Effect of MICS Strategy: MICS identifies a significantly higher proportion of effective reasoning paths than the vanilla method across the analyzed clinical scenarios.The analysis uses MICS-Score trends, with monotonically increasing or stable scores treated as effective.
- Effect of Training Set Settings: Adding medical VQA data improves VQA-RAD and SLAKE by average margins of 5% and 7.3%, respectively, but has limited impact on complex reasoning benchmarks.The training-set ablation evaluates whether foundational visual understanding data complements MMRP reasoning data.
- Effect of Training Strategy: The complete stage-wise curriculum outperforms partial strategies, while omitting Stage 3 particularly degrades reasoning and using only Stage 3 severely impairs VQA performance.On MMMU Health & Medicine, Stage 1, Stage 1+2, and Stage 3 alone trail the full model by 6.9%, 6.2%, and 3.5%, respectively.
5 Conclusion
The paper introduces MICS to construct high-quality multimodal medical reasoning data, then uses MMRP and stage-wise fine-tuning to train Chiron-o1. The resulting model achieves state-of-the-art benchmark performance and is presented as advancing medical CoT construction and medical MLLM reasoning.
- Conclusion: MICS preserves valid reasoning steps and eliminates incorrect ones, while MMRP supplies multi-task medical reasoning data with varying difficulty for Chiron-o1 training.The model is trained with a stage-wise fine-tuning approach.
A Deatails of MMRP
MMRP Part 1 generates filtered text-only medical QA data, while Part 2 aligns authentic medical-image analyses with images through coarse and precise correspondence rules.
- Part 1: Part 1 targets medical or clinical intern-level text-only question answering using cases with both medical imaging analysis C and case summary or discussion D.
- Part 1: 57,630 QA quadruplets remain in MMRP Part 1 after filtering incomplete cases, low-quality outputs, parsing failures, ambiguity, and hallucination-prone references.The initial collection contained 60,789 textual cases.
- Part 2: Part 2 contains approximately 3K annotated cases spanning 12 imaging modalities to align multimodal information from real medical images and their analyses.
- Part 2: Coarse alignment preserves the holistic correspondence between image descriptions and image order, whereas precise alignment maps text keywords to specific images.
- Part 2: Part 2 uses image descriptions directly sourced from authentic Radiopaedia medical imaging analyses rather than synthetically generated descriptions.
A.3 Part 3: Learning to Reason for Complex Problems via MICS!
MMRP Part 3 synthesizes complex multimodal clinical reasoning data from authentic case information and varied clinical question perspectives, with MICS producing step-by-step reasoning paths.
- A.3 Part 3: Learning to Reason for Complex Problems via MICS!: Part 3 uses broader authentic clinical information during CoT synthesis to reduce unfounded hallucinations rather than relying only on QA pairs.
- A.3 Part 3: Learning to Reason for Complex Problems via MICS!: Three clinical QA scenarios are designed with DeepSeek-V3, producing 8,328 complex visual QA pairs.
- A.3 Part 3: Learning to Reason for Complex Problems via MICS!: The scenarios include patient-to-doctor questions expressed colloquially and doctor-to-doctor questions focused on case condition, diagnosis, or treatment details.
- A.3 Part 3: Learning to Reason for Complex Problems via MICS!: MICS iteratively searches reasoning paths until maximum depth or a full-score path is reached, using three operations and configured mentor and intern models.
B Metrics
MICS-Score evaluates reasoning-path effectiveness using answer correctness and semantic similarity, while path-score trends compare search stability against a vanilla method.
- B Metrics: MICS-Score combines DeepSeek-V3 judgment of answer correctness with roberta-large semantic similarity between generated answers and ground truth.
- B Metrics: Path scores record the highest MICS-Score selected at each search iteration and are expected to increase or remain stable as reasoning deepens.
- B Metrics: Figure 7 qualitatively illustrates how MICS searches for effective medical reasoning paths.
- B Metrics: Figure 4(a) compares MICS with a vanilla method without MICS using monotonically increasing, non-increasing, constant, and fluctuating path-score trends.
C Training Sets
Chiron-o1 training combines standard medical VQA datasets with MMRP, and the reported ablations support the selected training-set configuration.
- C Training Sets: Chiron-o1 combines commonly used medical VQA datasets with MMRP during training to strengthen visual understanding, question answering, and reasoning foundations.
- C Training Sets: Ablation results in Table 3 indicate that the chosen training-set configuration is reasonable and effective.
- C Training Sets: Table 4 presents the distribution of datasets used at each training stage, including the Huatuo_PubMedVision alignment and instruction-tuning datasets.
D Experiments
Chiron-o1 is compared with Med-R1 and MedVLM-R1 across five MMMU(H&M) medical-question categories. It achieves substantial improvements across all categories, with especially large gains in clinically focused domains.
- MMMU(H&M) Results: Chiron-o1 achieves substantial performance improvements across all five MMMU(H&M) categories compared with Med-R1 and MedVLM-R1.The categories are Basic Medical Science, Clinical Medicine, Diagnostics and Laboratory Medicine, Pharmacy, and Public Health.
- MMMU(H&M) Results: The largest reported gains occur in Clinical Medicine, Diagnostics and Laboratory Medicine, and Pharmacy.These categories focus heavily on clinical problems.
- MMMU(H&M) Results: The authors attribute Chiron-o1’s superior clinical performance to MMRP’s construction from complex clinical cases.
E Limitations
The approach is constrained by the resource demands of MICS and by the current scale of the MMRP dataset. The authors identify reducing these costs and expanding the dataset as future priorities.
- Limitations: MICS requires collaboration between mentor and intern models, resulting in numerous costly API requests and substantial computational-resource demands.
- Limitations: MMRP requires further expansion in scale, which the authors identify as a focus of future work.
G Qualitative Analysis of Medical Reasoning Models
The qualitative analysis presents reasoning outputs from Med-R1, MedVLM-R1, and Chiron-o1, followed by six examples of Chiron-o1 reasoning. The supplied passages identify the examples but do not report their substantive reasoning outcomes.
- Model Reasoning Results: The qualitative analysis includes reasoning results for Med-R1, MedVLM-R1, and Chiron-o1.
- Chiron-o1 Examples: Six additional figures present Chiron-o1 reasoning on separate examples.