VLM2-Bench: A Closer Look at How Well VLMs Implicitly Link Explicit Matching Visual Cues

TL;DR

VLM2-Bench evaluates vision-language models' ability to link matching visual cues, revealing significant performance gaps and guiding future improvements in visual reasoning and training paradigms.

Contribution

Introduces VLM2-Bench, a comprehensive benchmark for assessing visual cue linking in VLMs, and provides insights into their limitations and directions for enhancement.

Findings

Models struggle with linking visual cues accurately.

Performance varies significantly across different models and prompting methods.

Recommendations for improving visual reasoning and training strategies.

Abstract

Visually linking matching cues is a crucial ability in daily life, such as identifying the same person in multiple photos based on their cues, even without knowing who they are. Despite the extensive knowledge that vision-language models (VLMs) possess, it remains largely unexplored whether they are capable of performing this fundamental task. To address this, we introduce \textbf{VLM2-Bench}, a benchmark designed to assess whether VLMs can Visually Link Matching cues, with 9 subtasks and over 3,000 test cases. Comprehensive evaluation across twelve VLMs, along with further analysis of various language-side and vision-side prompting methods, leads to a total of eight key findings. We identify critical challenges in models' ability to link visual cues, highlighting a significant performance gap. Based on these insights, we advocate for (i) enhancing core visual capabilities to improve adaptability and reduce reliance on prior knowledge, (ii) establishing clearer principles for integrating language-based reasoning in vision-centric tasks to prevent unnecessary biases, and (iii) shifting vision-text training paradigms toward fostering models' ability to independently structure and infer relationships among visual cues.