MediX-R1: Open Ended Medical Reinforcement Learning

Sahal Shaji Mullappilly; Mohammed Irfan Kurpath; Omair Mohamed; Mohamed Zidan; Fahad Khan; Salman Khan; Rao Anwer; Hisham Cholakkal

cs.CVFeb 26, 2026

MediX-R1: Open Ended Medical Reinforcement Learning

Authors: Sahal Shaji Mullappilly, Mohammed Irfan Kurpath, Omair Mohamed, Mohamed Zidan, Fahad Khan, Salman Khan, Rao Anwer, Hisham Cholakkal

Organizations: Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI) · Jubilee Mission Medical College and Research Institute · JJM Medical College

Abstract

We introduce MediX-R1, an open-ended Reinforcement Learning (RL) framework for medical multimodal large language models (MLLMs) that enables clinically grounded, free-form answers beyond multiple-choice formats. MediX-R1 fine-tunes a baseline vision-language backbone with Group Based RL and a composite reward tailored for medical reasoning: an LLM-based accuracy reward that judges semantic correctness with a strict YES/NO decision, a medical embedding-based semantic reward to capture paraphrases and terminology variants, and lightweight format and modality rewards that enforce interpretable reasoning and modality recognition. This multi-signal design provides stable, informative feedback for open-ended outputs where traditional verifiable or MCQ-only rewards fall short. To measure progress, we propose a unified evaluation framework for both text-only and image+text tasks that uses a Reference-based LLM-as-judge in place of brittle string-overlap metrics, capturing semantic correctness, reasoning, and contextual alignment. Despite using only $\sim51$ K instruction examples, MediX-R1 achieves excellent results across standard medical LLM (text-only) and VLM (image + text) benchmarks, outperforming strong open-source baselines and delivering particularly large gains on open-ended clinical tasks. Our results demonstrate that open-ended RL with comprehensive reward signals and LLM-based evaluation is a practical path toward reliable medical reasoning in multimodal models. Our trained models, curated datasets and source code are available at https://medix.cvmbzuai.com

Figures & tables

Model	Diverse Medical	Single-Stage	Interpretable	Open-Ended	Annotation-Free	Composite
	Modalities	RL	Reasoning	Responses	Reasoning	RL Reward
MedVLM-R1 ( Pan et al., 2025 )	✗	✓	✓	✗	✓	✗
BiMediX2 ( Mullappilly et al., 2024 )	✓	✗	✗	✓	✗	✗
HuatuoGPT-V ( Chen et al., 2024b )	✓	✗	✗	✓	✗	✗
MedGemma ( Sellergren et al., 2025 )	✓	✗	✓	✓	✗	✗
MedMO ( Deria et al., 2026 )	✓	✗	✓	✓	✗	✗

Table 1: Differences with existing Medical VLMs: MediX-R1 integrates diverse modalities, interpretable reasoning, and composite RL rewards, enabling practical clinical use.

Figure 2: MediX-R1: Overall Architecture The MediX-R1 reinforcement learning framework for open-ended medical reasoning. An input of a medical image and a natural language question is processed by MediX-R1. The model’s policy is trained using Group Based RL, which leverages a multi-faceted reward signal. This reward is composed of: a) an LLM-based reward for evaluating the overall quality and correctness of the output; b) an embedding-based reward to ensure semantic alignment; c) a format reward to enforce the desired output structure ( <think> and <answer> blocks); and d) a modality reward to ensure the response is grounded in the specified imaging modality. This reward-guided approach encourages the model to generate accurate and interpretable reasoning paths.

Figure 3

Benchmarks	MedVLM R1 2B	BiMediX2 8B	Huatuo GPT-V 7B	MedGemma 4B	MedGemma 27B	MedMO 8B	MediX-R1 2B	MediX-R1 8B	MediX-R1 30B
MMLU-Clinical	0.540	0.732	0.721	0.708	0.879	0.864	0.660	0.845	0.894
MMLU-Bio	0.549	0.792	0.708	0.706	0.972	0.951	0.806	0.951	0.993
MMLU-Med	0.451	0.694	0.653	0.605	0.866	0.827	0.699	0.879	0.890
MMLU-Genetics	0.560	0.790	0.710	0.820	0.940	0.900	0.680	0.900	0.980
MMLU-ProfMed	0.500	0.695	0.625	0.713	0.912	0.890	0.581	0.868	0.974
MMLU-Anat	0.519	0.659	0.600	0.556	0.793	0.785	0.563	0.763	0.874

Table 2: Evaluation Benchmark. The top section lists LLM (text-only) tasks and the bottom section lists VLM (image+text) tasks. Our three-stage evaluation setting evaluates both tasks in a unified framework. MediX-R1 achieves the highest average score across this diverse suite, demonstrating state-of-the-art performance among open models. Best and second best results are bold and underlined

Model	MMMU Medical Validation
MedVLM-R1 2B ( Pan et al., 2025 )	39.33
MedGemma 4B ( Sellergren et al., 2025 )	50.00
HuatuoGPT-V 7B ( Chen et al., 2024b )	47.33
BiMediX2 8B ( Mullappilly et al., 2024 )	39.33
Qwen3-VL 8B ( Yang et al., 2025 )	62.66
MedMO 8B ( Deria et al., 2026 )	62.66

Table 3: MMMU Medical Val results. Accuracy on the Health & Medical validation subset of MMMU (image-text).

Panel of Figure 6 — Figure 6: Qualitative examples of MediX-R1 . (Top, Microscopy) Correctly identifies the optic tract in section G with interpretable reasoning. (Bottom, X-ray) Explains why heart size appears smaller in PA vs. AP view. MediX-R1 generates clinically grounded, open-ended answers across modalities.

Evaluations	Default	Embedding	LLM	LLM +	MediX-R1
	Reward	Reward	Reward	Embedding
LLM Tasks	0.660	0.640	0.666	0.686	0.687
VLM Tasks	0.382	0.409	0.400	0.410	0.431
Overall AVG	0.562	0.558	0.572	0.589	0.597

Table 4: Reward ablation across validation benchmarks. Single signals like default string matching, embedding-only or LLM-only are weaker. Combining LLM + embedding improves robustness, and the MediX-R1 composite (LLM accuracy + embedding-based semantics + modality recognition) yields the best overall average.

Model	Baseline	+ Composite Rewards
SmolVLM2-2.2B ( Marafioti et al., 2025 )	0.410	0.432
Qwen3-VL-2B ( Yang et al., 2025 )	0.529	0.554
Qwen3-VL-8B ( Yang et al., 2025 )	0.666	0.688
Qwen3-VL-30B ( Yang et al., 2025 )	0.698	0.736

Table 5: Baseline comparison across backbones (overall AVG). “Baseline” is the original backbone; “+Composite Rewards” applies our RL with composite rewards on the same backbone.

Evaluations	Baseline	GRPO	GSPO	DAPO
LLM Evaluations (text only)	0.675	0.687	0.689	0.701
VLM Evaluations (image + text)	0.376	0.431	0.439	0.445
Overall AVG	0.570	0.597	0.600	0.610

Table 6: RL Method ablation across benchmarks. Using the same composite reward, different RL algorithms (GRPO/GSPO/DAPO) consistently outperform the baseline across both LLM and VLM tasks, with DAPO achieving the highest overall average.

Figure 7: Overall validation reward vs training step across reward designs . Training with individual signals LLM-only or embedding-only shows volatility and reward hacking, while LLM+embedding reduces but does not eliminate instability. MediX-R1 uses a composite reward which stabilizes learning and delivers the highest final reward and best overall performance.

Model	Score (%)
MedVLM-R1	27.57
MedGemma	43.18
LLaVA-Med	44.29
BiMediX2	46.51
HuatuoGPT	48.81
MediX-R1 (Ours)	51.11

Table 7: Performance comparison on the MedPix 2.0 dataset.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Samples
PMC_VQA_SUBSET	25000
SLAKE	4919
RAD_VQA	1793
PATH	19623
Total	51335

Appendix

Table 8: Modality Breakdown and Source Dataset composition

Figure 8: MediX-R1 - Report Generation: Case 2

Figure 9: MediX-R1 - Report Generation: Case 1

Explore similar work

OpenMedReason: Scientific Reasoning Supervision for Medical Vision-Language Models
Jun 10, 2026Negin Baghbanzadeh, Pritam Sarkar, Michael Colacci +6Medical VQA Medical VLMs
York University · Vector Institute · University of British Columbia +5
Lingshu: A Generalist Foundation Model for Unified Multimodal Medical Understanding and Reasoning
Jun 8, 2025Weiwen Xu, Hou Pong Chan, Long Li +15Medical Report Generation Multimodal Pretraining
LASA Team · DAMO Academy, Alibaba Group
Fathom-Vaidya: Advancing Medical Reasoning with Rubric-Based Rewards
Sep 21, 2026Kalash Shah, Kunal Singh, Snehan J +1RL for Language Model Reasoning Rubric-Based RL
Fractal AI Research

Jun 10, 2026J/K move · Enter open · S save

OpenMedReason: Scientific Reasoning Supervision for Medical Vision-Language Models

Negin Baghbanzadeh, Pritam Sarkar, Michael Colacci, Abeer Badawi+5

York University · Vector Institute · University of British Columbia+5

Medical VQA Medical VLMs VLM Reasoning Visually Grounded Reasoning Medical Image Benchmarks

Open paper AskPDF arXiv

High-stakes clinical use of large vision-language models (LVLMs) requires reasoning that is grounded in visual evidence and clinical knowledge, not just correct final answers. We introduce OpenMedReason, a large-scale, open multimodal medical reasoning corpus comprising approximately 450K image-question-answer instances whose reasoning traces are primarily derived from curated biomedical, human-authored scientific articles. OpenMedReason provides high-fidelity supervision beyond synthetic chains of thought, covering diverse medical domain vision modalities such as radiological scans, microscopic images, visible light photographs, charts, and others. We complement it with OpenMedReason-Bench, a held-out benchmark that allows fine-grained evaluation of LVLMs along three complementary axes of capability, including perception, medical knowledge, and rationale, enabling diagnostic evaluation beyond final-answer accuracy. OpenMedReason is a rich training resource that exhibits its effectiveness in both supervised fine-tuning (SFT) and reinforcement-based alignment. Training with OpenMedReason yields a 20% average improvement in VQA accuracy over the base model and achieves performance within 4.2% of the strongest comparable-scale medical LVLMs. Fine-grained performance analysis confirms that the gains are not concentrated in any single axis: OpenMedReason improves perception, medical knowledge, and rationale jointly, and its reasoning traces are preferred over those of the base model in 86.1% of pairwise comparisons. We release the code and dataset at huggingface.co/datasets/neginb/OpenMedReason.

Abstract

Explore similar work

OpenMedReason: Scientific Reasoning Supervision for Medical Vision-Language Models

Lingshu: A Generalist Foundation Model for Unified Multimodal Medical Understanding and Reasoning

Fathom-Vaidya: Advancing Medical Reasoning with Rubric-Based Rewards