Benchmark Success, Clinical Failure: When Reinforcement Learning Optimizes for Benchmarks, Not Patients
On the map: LLM reasoning → Clinical and biomedical applications
Abstract (arXiv)
Recent Reinforcement Learning (RL) advances for Large Language Models (LLMs) have improved reasoning tasks, yet their resource-constrained application to medical imaging remains underexplored. We introduce ChexReason, a vision-language model trained via R1-style methodology (SFT followed by GRPO) using only 2,000 SFT samples, 1,000 RL samples, and a single A100 GPU. Evaluations on CheXpert and NIH benchmarks reveal a fundamental tension: GRPO recovers in-distribution performance (23% improvement on CheXpert, macro-F1 = 0.346) but degrades cross-dataset transferability (19% drop on NIH). This mirrors high-resource models like NV-Reason-CXR-3B, suggesting the issue stems from the RL paradigm rather than scale. We identify a generalization paradox where the SFT checkpoint uniquely improves on NIH before optimization, indicating teacher-guided reasoning captures more institution-agnostic features. Furthermore, cross-model comparisons show structured reasoning scaffolds benefit general-purpose VLMs but offer minimal gain for medically pre-trained models. Consequently, curated supervised fine-tuning may outperform aggressive RL for clinical deployment requiring robustness across diverse populations.
Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Clinical and biomedical applications, one of 17 directions in LLM reasoning. Builds on 30 of its 40 references that are in the map. First posted on arXiv 28 Dec 2025.
Builds on
30 of 40 references.
Closest ancestors · 10
Same direction, newest first
- 2025Reasoning Visual Language Model for Chest X-Ray Analysis28 Oct
- 2025X-Ray-CoT: Interpretable Chest X-ray Diagnosis with Vision-Language Models via Chain-of-Thought Reasoning17 Aug
- 2025GEMeX-RMCoT: An Enhanced Med-VQA Dataset for Region-Aware Multimodal Chain-of-Thought Reasoning22 Jun
- 2025RARL: Improving Medical VLM Reasoning and Generalization with Reinforcement Learning and LoRA under Data and Hardware Constraints7 Jun
- 2025Toward Effective Reinforcement Learning Fine-Tuning for Medical VQA in Vision-Language Models20 May
- 2025ChestX-Reasoner: Advancing Radiology Foundation Models with Reasoning through Step-by-Step Verification29 Apr
- 2025GMAI-VL-R1: Harnessing Reinforcement Learning for Multimodal Medical Reasoning2 Apr
- 2025Med-R1: Reinforcement Learning for Generalizable Medical Reasoning in Vision-Language Models18 Mar
- 2025Med-RLVR: Emerging Medical Reasoning from a 3B base model via reinforcement Learning27 Feb
- 2025MedVLM-R1: Incentivizing Medical Reasoning Capability of Vision-Language Models (VLMs) via Reinforcement Learning26 Feb
Other references · 20
- 2021LoRA: Low-Rank Adaptation of Large Language Modelscited by 14,084
- 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 8,851
- 2019CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparisoncited by 1,410
- 2025Understanding R1-Zero-Like Training: A Critical Perspectivecited by 1,276
- 2019MIMIC-CXR-JPG, a large publicly available database of labeled chest radiographscited by 590
- 2025MedRAX: Medical Reasoning Agent for Chest X-raycited by 68
- 2025Vision-R1: Evolving Human-Free Alignment in Large Vision-Language Models via Vision-Guided Reinforcement Learningcited by 68
- 2024CXPMRG-Bench: Pre-training and Benchmarking for X-ray Medical Report Generation on CheXpert Plus Datasetcited by 33
- 2024ReXrank: A Public Leaderboard for AI-Powered Radiology Report Generationcited by 24
- 2024MMedPO: Aligning Medical Vision-Language Models with Clinical-Aware Multimodal Preference Optimizationcited by 21
- 2023When More is Less: Incorporating Additional Datasets Can Hurt Performance By Introducing Spurious Correlationscited by 18
- 2023Interpretable Medical Image Visual Question Answering via Multi-Modal Relationship Graph Learningcited by 16
8 more
- 2022Significantly improving zero-shot X-ray pathology classification via fine-tuning pre-trained image-text encoderscited by 15
- 2023Beyond Distribution Shift: Spurious Features Through the Lens of Training Dynamicscited by 11
- 2023Robust Fine-Tuning of Vision-Language Models for Domain Generalizationcited by 7
- 2024Direct Preference Optimization for Suppressing Hallucinated Prior Exams in Radiology Report Generationcited by 7
- 2025Boosting multi-demographic federated learning for chest radiograph analysis using general-purpose self-supervised representationscited by 6
- 2024LiteGPT: Large Vision-Language Model for Joint Chest X-ray Localization and Classification Taskcited by 4
- 2025Preventing Shortcut Learning in Medical Image Analysis through Intermediate Layer Knowledge Distillation from Specialist Teacherscited by 4
- 2025Enhancing zero-shot learning in medical imaging: integrating clip with advanced techniques for improved chest x-ray analysiscited by 2
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2025MedReason-R1: Learning to Reason for CT Diagnosis with Reinforcement Learning and Local Zoom
- 2025Adaptive Reinforcement for Open-ended Medical Reasoning via Semantic-Guided Reward Collapse Mitigation
- 2026MMedExpert-R1: Strengthening Multimodal Medical Reasoning via Domain-Specific Adaptation and Clinical Guideline Reinforcement
- 2025MedReasoner: Reinforcement Learning Drives Reasoning Grounding from Clinical Thought to Pixel-Level Precision
- 2026Breaking Failure Cascades: Step-Aware Reinforcement Learning for Medical Multimodal Reasoning
- 2025Enhancing Radiology Report Generation and Visual Grounding using Reinforcement Learning
- 2026MediX-R1: Open Ended Medical Reinforcement Learning
- 2026Regulating Anatomy-Aware Rewards via Trajectory-Integral Feedback for Volumetric Computed Tomography Analysis