ReXInTheWild: A Unified Benchmark for Medical Photograph Understanding

2026-03-19

AI Summary

This work addresses the lack of comprehensive evaluation benchmarks for visual-language models in understanding real-world, everyday medical photographs—a task requiring both natural image perception and clinical reasoning. To bridge this gap, the authors introduce the first fine-grained, clinically grounded multimodal benchmark derived from authentic medical photography, comprising 484 literature-sourced images and 955 expert-validated multiple-choice questions spanning seven clinical domains. Systematic evaluation using this benchmark reveals that general-purpose models (e.g., Gemini-1.5 achieving 78% accuracy) substantially outperform specialized medical models such as MedGemma (37% accuracy). The study further identifies four characteristic error patterns, offering clear directions for future model refinement and development in medical multimodal understanding.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Multi-modal Vision
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data

Abstract

Everyday photographs taken with ordinary cameras are already widely used in telemedicine and other online health conversations, yet no comprehensive benchmark evaluates whether vision-language models can interpret their medical content. Analyzing these images requires both fine-grained natural image understanding and domain-specific medical reasoning, a combination that challenges both general-purpose and specialized models. We introduce ReXInTheWild, a benchmark of 955 clinician-verified multiple-choice questions spanning seven clinical topics across 484 photographs sourced from the biomedical literature. When evaluated on ReXInTheWild, leading multimodal large language models show substantial performance variation: Gemini-3 achieves 78% accuracy, followed by Claude Opus 4.5 (72%) and GPT-5 (68%), while the medical specialist model MedGemma reaches only 37%. A systematic error analysis also reveals four categories of common errors, ranging from low-level geometric errors to high-level reasoning failures and requiring different mitigation strategies. ReXInTheWild provides a challenging, clinically grounded benchmark at the intersection of natural image understanding and medical reasoning. The dataset is available on HuggingFace.

Problem

• medical photograph understanding
• vision-language models
• benchmark
• multimodal large language models
• medical reasoning

Innovation

• medical photograph understanding
• vision-language models
• multimodal benchmark
• clinical reasoning
• error analysis