Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Berger, Armin, Bergau, Manuela, Schneider, Helen, Ahmad, Saad, Lagones, Tom Anglim, Brugnara, Gianluca, Foltyn-Dumitru, Martha, Schlamp, Kai, Vollmuth, Philipp, Sifa, Rafet
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2512.23090
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866908743198834688
author Berger, Armin
Bergau, Manuela
Schneider, Helen
Ahmad, Saad
Lagones, Tom Anglim
Brugnara, Gianluca
Foltyn-Dumitru, Martha
Schlamp, Kai
Vollmuth, Philipp
Sifa, Rafet
author_facet Berger, Armin
Bergau, Manuela
Schneider, Helen
Ahmad, Saad
Lagones, Tom Anglim
Brugnara, Gianluca
Foltyn-Dumitru, Martha
Schlamp, Kai
Vollmuth, Philipp
Sifa, Rafet
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2512_23090
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Benchmark Success, Clinical Failure: When Reinforcement Learning Optimizes for Benchmarks, Not Patients
Berger, Armin
Bergau, Manuela
Schneider, Helen
Ahmad, Saad
Lagones, Tom Anglim
Brugnara, Gianluca
Foltyn-Dumitru, Martha
Schlamp, Kai
Vollmuth, Philipp
Sifa, Rafet
Artificial Intelligence
Machine Learning
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.
title Benchmark Success, Clinical Failure: When Reinforcement Learning Optimizes for Benchmarks, Not Patients
topic Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2512.23090