VLM-SubtleBench: How Far Are VLMs from Human-Level Subtle Comparative Reasoning?

TL;DR

VLM-SubtleBench is a new benchmark for evaluating vision-language models on subtle, real-world image differences across various domains, revealing significant gaps compared to human reasoning.

Contribution

The paper introduces VLM-SubtleBench, a comprehensive benchmark for subtle comparative reasoning in diverse domains, and provides extensive evaluations exposing current model limitations.

Findings

VLMs underperform humans on subtle differences

Performance varies significantly across difference types and domains

Identifies key areas for improving VLM reasoning capabilities

Abstract

The ability to distinguish subtle differences between visually similar images is essential for diverse domains such as industrial anomaly detection, medical imaging, and aerial surveillance. While comparative reasoning benchmarks for vision-language models (VLMs) have recently emerged, they primarily focus on images with large, salient differences and fail to capture the nuanced reasoning required for real-world applications. In this work, we introduce VLM-SubtleBench, a benchmark designed to evaluate VLMs on subtle comparative reasoning. Our benchmark covers ten difference types - Attribute, State, Emotion, Temporal, Spatial, Existence, Quantity, Quality, Viewpoint, and Action - and curate paired question-image sets reflecting these fine-grained variations. Unlike prior benchmarks restricted to natural image datasets, our benchmark spans diverse domains, including industrial, aerial, and medical imagery. Through extensive evaluation of both proprietary and open-source VLMs, we reveal systematic gaps between model and human performance across difference types and domains, and provide controlled analyses highlighting where VLMs' reasoning sharply deteriorates. Together, our benchmark and findings establish a foundation for advancing VLMs toward human-level comparative reasoning.