Research map / Vision-language and multimodal / Multimodal LLMs

OpenMedReason: Scientific Reasoning Supervision for Medical Vision–Language Models

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 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.

Introduces OpenMedReason, a ~450K medical VQA reasoning corpus with source-grounded traces from biomedical article figures, plus a capability-decomposed benchmark, and shows SFT→GRPO on Qwen2.5-VL-7B improves medical VQA.

Problem

Medical LVLMs need reasoning supervision grounded in visual evidence and clinical knowledge, but existing open resources either lack reasoning traces or rely on end-to-end LLM rollouts, and evaluation focuses on final-answer accuracy.

Approach

Derive reasoning traces from curated, human-authored biomedical figure captions and in-text context (Open-PMC) rather than purely synthetic LLM rollouts, structure them as modality→visual evidence→clinical interpretation→answer, and evaluate with a checklist-based trace score decomposed into perception, knowledge, and rationale axes.

Key results

• SFT+GRPO on OpenMedReason raises Qwen2.5-VL-7B average VQA accuracy from 53.10 to 60.04 across the reported VQA benchmarks (cumulative gain ~7 points), with 78.51 on OpenMedReason-Bench vs 47.11 for base.
• On OpenMedReason-Bench trace evaluation, the model reaches a Reasoning Trace Score of 77.8 vs 70.9 for gpt-5-mini and 37.4 for the Qwen2.5-VL-7B base.
• Classification averages rise from 43.12 (base) to 65.07, above Lingshu 7B (61.76) on the reported classification benchmarks.

Surprising findings

• Adding SFT on OpenMedReason traces without Open-PMC data gives in-distribution gains but the Open-PMC-derived subset is what drives the out-of-distribution improvement.
• Pass@10 gains on MedXpertQA and JAMA exceed Pass@1 gains, suggesting expanded reachable solution coverage.

Limitations

• Derived from published biomedical figures, inheriting literature biases such as over-representation of rare cases and uneven modality coverage.
• Benchmark scores and traces are not evidence of clinical safety or deployment readiness.
• LoRA training was used due to computational constraints, and the vision tower was frozen.

Experimental setup

Datasets: Open-PMC-18M SLAKE VQA-RAD PathVQA PMC-VQA MedXpertQA JAMA CC-MM HAM10000 EyePACS HyperKvasir BrainTumorMRI VinDr-CXR

Benchmarks: OpenMedReason-Bench SLAKE VQA-RAD PathVQA PMC-VQA MedXpertQA-MM JAMA CC-MM HAM10000 EyePACS HyperKvasir BrainTumorMRI VinDr-CXR

Metrics: accuracy presence correctness reasoning trace score pass@1 pass@10 pairwise win rate

Baselines: Lingshu 7B OctoMed 7B MedGemma 27B MedGemma 1.5 4B QoQ-Med-VL 7B MedVL-Thinker 7B Qwen2.5-VL-7B gpt-5-mini claude-haiku-4.5 gemini-3-flash

Builds on: GRPO Open-PMC-18M Qwen2.5-VL

Closest papers by meaning

• OctoMed: Data Recipes for State-of-the-Art Multimodal Medical Reasoning (CVPR 2026 · similarity 0.91)
• Med-CMR: A Fine-Grained Benchmark Integrating Visual Evidence and Clinical Logic for Medical Complex Multimodal Reasoning (CVPR 2026 · similarity 0.91)
• Medical thinking with multiple images (ICLR 2026 · similarity 0.91)
• OmniMedVQA: A New Large-Scale Comprehensive Evaluation Benchmark for Medical LVLM (CVPR 2024 · similarity 0.90)
• Visual and Domain Knowledge for Professional-level Graph-of-Thought Medical Reasoning (ICML 2025 · similarity 0.90)
• MedReasoner: Reinforcement Learning Drives Reasoning Grounding from Clinical Thought to Pixel-Level Precision (AAAI 2026 · similarity 0.90)