Multimodal Large Language Models (MLLMs) have achieved remarkable success in visual and textual tasks. However, their visual relational reasoning capabilities remain insufficiently explored. Further more, most real-world Internet images used in model training lead models to guess answers via latent bias before actual reasoning. This highlights the need for further investigation into how MLLMs reason about relationships in abstract scenes. Recent studies suggest that visual relational reasoning is closely tied to intelligence, inspired by this, we introduce VRR-BENCH, a benchmark designed to evaluate the Visual Relational Reasoning abilities of MLLMs. The dataset is divided into Non-Relational and Visual Relational reasoning tasks across three levels, each involving different combinations of object attributes. Our evaluation of six MLLMs, including GPT-4o, reveals that relational tasks are generally more challenging, with an average accuracy drop of 22.01%. Additionally, model performance varied with different numbers of object attribute combinations, indicating diverse challenges across tasks. We conducted comprehensive and progressive testing using VRR-BENCH, and we believe this research can serve as a reference for future work.