InfiMed: Low-Resource Medical MLLMs with Advancing Understanding and Reasoning
InfiMed integrates reflective-patterned CoT and rule-based RLVR, boosting low-resource medical multimodal model performance, achieving 7 SOTA benchmarks.
Key Findings
Methodology
This work introduces a framework combining high-quality textual reasoning data with multimodal medical datasets during supervised fine-tuning (SFT). It employs reflective-pattern injected chain-of-thought (CoT) to address data sparsity, generating diverse reasoning trajectories. Response quality is evaluated via rejection sampling, and rule-based rewards (e.g., Jaccard similarity) guide reinforcement learning (RLVR). The core algorithms include Qwen2.5-VL-72B as the generator, with Group Relative Policy Optimization (GRPO) for policy updates, optimizing for format and accuracy rewards, thus enhancing reasoning depth and robustness under limited data conditions.
Key Results
- InfiMed-RL-3B achieves an average accuracy of 59.2% across seven medical benchmarks, surpassing InternVL3-8B (57.3%) and MedGemma-4B-IT (54.8%) with similar parameter counts, demonstrating high efficiency.
- Training on 188K SFT samples and 36K RLVR samples, the model exhibits strong reasoning, comprehension, and self-correction capabilities, especially in sparse data scenarios.
- Incorporating reflective-patterned CoT significantly improves performance on complex reasoning tasks, validating the effectiveness of the reflection mechanism in medical contexts.
Significance
This research addresses the critical challenge of low-resource medical AI by proposing a novel training paradigm that combines reflective reasoning with reinforcement learning. It enhances model interpretability, reliability, and reasoning depth, which are essential for clinical trustworthiness. The approach reduces dependency on large datasets, making high-performance medical AI accessible to resource-constrained settings. The framework paves the way for more explainable and robust AI systems in healthcare, facilitating safer deployment and broader adoption in real-world clinical environments.
Technical Contribution
The study introduces a pioneering integration of reflective-patterned CoT with rule-based RLVR, enabling models to self-evaluate and correct errors during training. The rejection sampling combined with rule rewards ensures high-quality response generation, guiding exploration toward deeper reasoning paths. The methodology achieves state-of-the-art results with significantly fewer parameters and training data, offering a new theoretical and engineering paradigm for low-resource medical multimodal AI. This work also provides insights into the design of self-reflective reasoning mechanisms and their impact on model robustness.
Novelty
This is the first work to embed reflective-patterned CoT into low-resource medical multimodal models, leveraging rejection sampling and rule-based rewards for reinforcement learning. Unlike traditional fine-tuning that relies solely on large datasets, this approach enhances reasoning depth and self-correction capabilities with limited data. It introduces a novel training pipeline that combines self-reflection, data augmentation, and exploration strategies, setting a new standard for efficient, interpretable medical AI models.
Limitations
- Despite improvements, the model still struggles with extremely sparse or noisy data, leading to occasional reasoning errors. Its generalization to unseen or highly complex cases remains limited.
- The training process demands substantial computational resources, especially for generating and evaluating diverse responses, which may hinder scalability.
- The current reward functions may not fully capture clinical correctness or explainability, necessitating further refinement for real-world deployment.
Future Work
Future directions include developing adaptive reflection mechanisms that dynamically adjust reasoning strategies based on task complexity, integrating more diverse clinical datasets, and exploring multi-task learning to enhance generalization. Additionally, efforts will focus on reducing computational costs and improving explainability to facilitate clinical adoption. Extending the framework to real-world deployment scenarios and validating in clinical settings will be crucial steps forward.
AI Executive Summary
The rapid evolution of multimodal large language models (MLLMs) has revolutionized AI applications across various domains, notably in visual understanding and mathematical reasoning. However, their deployment in medicine faces significant hurdles, primarily due to the scarcity of high-quality, annotated multimodal datasets. Existing models often rely on large-scale data and parameter sizes, which are resource-intensive and less feasible for low-resource settings. Addressing this gap, the present study introduces InfiMed, a series of models designed to excel in low-resource medical scenarios.
InfiMed's core innovation lies in combining high-quality textual reasoning data with multimodal medical data during supervised fine-tuning (SFT). To mitigate data sparsity, the authors synthesize reflective-patterned chain-of-thought (CoT) data, enabling the model to develop initial reflective reasoning capabilities. This structured reasoning foundation is further enhanced through reinforcement learning with verifiable rewards (RLVR), which employs rejection sampling and rule-based reward functions—such as Jaccard similarity—to guide exploration. The models are trained with 188K SFT samples and 36K RLVR samples, demonstrating significant performance gains.
Experimental results across seven diverse medical benchmarks show that InfiMed-RL-3B achieves an average accuracy of 59.2%, outperforming larger models like InternVL3-8B. This indicates that the combined approach of reflective reasoning and reinforcement learning effectively enhances the model's inference depth, robustness, and clinical relevance. The framework not only advances the state-of-the-art but also offers a scalable, resource-efficient pathway for deploying high-performance medical AI in resource-constrained environments.
The significance of this work extends beyond performance metrics. It introduces a paradigm shift towards more interpretable, self-correcting AI systems capable of reasoning akin to human clinicians. Such models can support medical decision-making, diagnostics, and report generation with increased trustworthiness. Despite these advances, challenges remain, including handling extreme data sparsity and computational costs. Future research will focus on adaptive reflection strategies, broader clinical validation, and real-world deployment, aiming to make AI-driven healthcare more accessible, reliable, and explainable.
Deep Analysis
Background
Medical AI的发展经历了从规则基础到深度学习的跃迁。早期依赖专家规则，难以应对复杂场景。近年来，代表性模型如HuatuoGPT、Med-PaLM在医学问答和报告生成中表现出色，但多模态融合仍受数据限制。大模型如GPT-4、GPT-5在通用任务中表现优异，但在医学场景中受限于数据稀缺和推理深度不足。现有数据集如PMC-15M、VQA-RAD、SLAKE逐步丰富，但信息稀疏、标注不足，限制模型推理能力。整体来看，医学AI面临数据不足、推理深度有限和可解释性差的挑战，亟需创新训练策略和模型架构突破瓶颈。
Core Problem
当前医学多模态模型普遍存在数据稀疏、信息不足的问题，导致推理深度受限，难以满足临床复杂场景需求。传统微调依赖大量标注，成本高且难以扩展。强化学习虽能提升探索能力，但在医学中的应用有限，缺乏有效奖励机制和反思能力，限制模型推理深度和可靠性。如何在有限资源下增强模型推理、理解和自我校正能力，成为核心难题。特别是在信息稀疏场景中，模型易陷入浅层记忆，影响诊断的准确性和可信度。
Innovation
本文提出结合反思模式注入的链式推理（CoT）与基于规则的强化学习（RLVR），创新性缓解数据稀疏带来的推理不足。首先，利用高质量文本推理与多模态医学数据进行微调，恢复基础推理能力。其次，设计反思样式的CoT，生成多样推理轨迹，提升自我校正和错误检测。再次，采用拒绝采样筛选响应，结合规则奖励（如Jaccard相似度）引导探索，增强推理深度。最后，提出InfiMed模型，在参数有限条件下实现多项基准SOTA，验证了方法有效性。这一框架突破了传统微调对大数据的依赖，为低资源医学AI提供新思路。
Methodology
- �� 结合高质量文本推理数据与多模态医学数据进行监督微调（SFT），强化基础推理能力。• 引入反思样式的链式思考（CoT），通过生成多样推理轨迹，提升自我校正能力。• 利用拒绝采样筛选响应，将高质量响应与错误响应结合，丰富训练样本。• 设计基于规则的奖励函数（如Jaccard相似度）评估响应的正确性和格式，指导强化学习（RLVR）。• 采用GRPO策略优化，最大化优势函数，确保训练稳定。• 在训练中结合格式与准确度奖励，逐步提升推理深度和鲁棒性。
Experiments
采用七项医学多模态基准（如MMMU-H&M、VQA-RAD、SLAKE等），对比InfiMed系列与其他模型（如GPT-4、MedGemma等）。训练数据包括188K SFT样本与36K RLVR样本，参数规模为3B。指标为准确率，模型在不同任务中表现优异。通过消融实验验证反思机制和奖励设计的贡献，分析模型在稀疏信息场景中的表现差异。实验还评估模型推理深度、可解释性和鲁棒性，确保实际应用可靠。
Results
InfiMed-RL-3B在七项基准中平均准确率达59.2%，优于同规模模型InternVL3-8B（57.3%）和MedGemma-4B-IT（54.8%）。引入反思样式的CoT显著提升复杂推理任务表现，验证反思机制有效。模型在稀疏信息场景中的推理深度增强，展现良好泛化能力。实验结果证明，结合反思与强化学习的训练策略在低资源条件下实现性能突破，为医学AI提供新范式。
Applications
模型可应用于临床辅助诊断、医学报告自动生成、医学影像分析，特别适合资源有限的医疗环境。通过提升推理深度和解释性，帮助医生做出更准确判断。未来结合电子健康记录、多模态影像等多源数据，拓展应用场景，推动智慧医疗发展。
Limitations & Outlook
模型在极端稀疏或噪声较多的医学数据中仍存在推理误差，反思机制泛化能力有限。训练过程对硬件资源需求较高，稳定性和可扩展性待提升。未来需优化反思样式生成策略，增强模型适应性和鲁棒性。
Plain Language Accessible to non-experts
想象你在一家工厂工作，工厂里有很多不同的机器和流程。每台机器都需要按照一定步骤操作，才能生产出合格的产品。如果一台机器出错，你会自己反思，找出问题所在，然后修理它。这个工厂的目标是让机器自己检查和修正错误，确保生产顺利。类似地，医学AI模型也是这样：它们要理解各种医学图片和文字，自己判断是否合理，然后不断学习改进。引入反思机制就像让机器在出错时自己反思，找到问题，改正答案。这样，即使数据不多，它们也能像经验丰富的工人一样，做出准确判断，帮助医生更好地诊断疾病。
ELI14 Explained like you're 14
想象你在学校学习，老师布置了很多作业。有时候题目很难，你会遇到不会的地方。为了弄懂，你会反复思考，甚至试用不同的方法。有时候你会发现自己之前的方法不对，然后反思改正。医学AI模型也是一样：它们要理解医学图片和文字，回答复杂问题。有时候信息不完整或模糊，模型需要像你一样反思，找到问题，改正答案。作者设计了一种让模型自己反思的办法，就像你在学习中不断总结经验。这样，模型就能变得更聪明、更可靠，帮助医生做出更准确的判断。这个方法让AI在有限数据下，也能学会像专家一样推理和反思，变得更有用。
Abstract
Multimodal Large Language Models (MLLMs) have achieved remarkable progress in domains such as visual understanding and mathematical reasoning. However, their application in the medical domain is constrained by two key challenges: (1) multimodal medical datasets are scarce and often contain sparse information, limiting reasoning depth; and (2) Reinforcement Learning with Verifiable Rewards (RLVR), though effective in general domains, cannot reliably improve model performance in the medical domain. To overcome these challenges, during the supervised fine-tuning (SFT) stage, we incorporate high-quality textual reasoning data and general multimodal data alongside multimodal medical data to efficiently enhance foundational medical capabilities and restore the base model's reasoning ability. Moreover, considering that there are some multimodal medical datasets with sparse information, we further synthesize reflective-pattern-injected chain-of-thought (CoT) in addition to general CoT samples, equipping the model with initial reflective reasoning capabilities that provide a structured foundation for subsequent RLVR training. Finally, we introduce our InfiMed-Series models, InfiMed-SFT-3B and InfiMed-RL-3B, both of which deliver state-of-the-art performance across seven multimodal medical benchmarks. Notably, InfiMed-RL-3B achieves an average accuracy of 59.2%, outperforming even larger models like InternVL3-8B, which achieves 57.3%. Specifically, during the SFT phase, we utilized 188K samples, while the RLVR phase incorporated 36K samples, demonstrating the efficacy of both training strategies in achieving superior performance. We also conducted a series of extensive experiments, which provide valuable insights that contribute to advancing the performance of MLLMs in medical scenarios.