MedLayBench-V

A Large-Scale Benchmark for Expert–Lay Semantic Alignment in Medical Vision-Language Models

Han Jang^*, Junhyeok Lee^*, Heeseong Eum, Kyu Sung Choi^†

^* Equal contribution. ^† Corresponding author.

Seoul National University · SNU College of Medicine · SNU Hospital · AICON Lab

ACL 2026 Findings Oral

arXiv

Code 🤗 Dataset

TL;DR

80K medical images, paired with expert and verified lay captions.

Abstract

Medical Vision-Language Models (Med-VLMs) have achieved expert-level proficiency in interpreting diagnostic imaging. However, current models are predominantly trained on professional literature, limiting their ability to communicate findings in the lay register required for patient-centered care. To bridge this gap we introduce MedLayBench-V, the first large-scale multimodal benchmark dedicated to expert–lay semantic alignment. Unlike naive simplification approaches that risk hallucination, our dataset is constructed via a Structured Concept-Grounded Refinement (SCGR) pipeline. SCGR enforces strict semantic equivalence by integrating UMLS Concept Unique Identifiers (CUIs) with micro-level entity constraints. MedLayBench-V provides a verified foundation for training and evaluating next-generation Med-VLMs capable of bridging the communication divide between clinical experts and patients.

The SCGR Pipeline

SCGR Pipeline overview — **Overview of SCGR.** (a) Expert input. (b) Concept-Knowledge Alignment couples UMLS CUI mapping with SciSpacy NER to build a constraint set C; MedlinePlus definitions seed a lay draft, then Llama-3.1-8B refines for fluency under strict factual constraints. (c) Lay output is clinically accurate and patient-accessible.

BibTeX

@misc{jang2026medlaybenchvlargescalebenchmarkexpertlay,
  title     = {MedLayBench-V: A Large-Scale Benchmark for Expert-Lay Semantic Alignment in Medical Vision Language Models},
  author    = {Jang, Han and Lee, Junhyeok and Eum, Heeseong and Choi, Kyu Sung},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year      = {2026}
}