Investigating Relational Reasoning in VLMs
2026-08-24 • Computer Vision and Pattern Recognition
Computer Vision and Pattern Recognition
AI summaryⓘ
The authors studied a modern vision-language model called Qwen3-VL-4B to see if it truly understands relationships in images or just relies on language hints. They created a special set of simple shape images and questions to carefully test this. Their experiments show that the model uses some real visual reasoning but also depends a lot on language shortcuts. This means the model’s understanding of images is mixed between seeing and guessing based on words.
Vision-Language ModelsVisual reasoningQwen3-VL-4BSynthetic datasetLanguage cuesVisual relationsShortcut strategiesCausal reliance
Authors
Adhithya Laxman Ravi Shankar Geetha, Aulia Kharis Rakhmasari, Haleema Ramzan, Xander Yap
Abstract
Vision-Language Models (VLMs) achieve strong performance in visual reasoning tasks, but it remains unclear whether they understand visual relations, or simply employ shortcuts such as language cues or priors. To investigate this, we use the Qwen3-VL-4B (Bai et al., 2025), a modern VLM, to decode how visual information is encoded across depths. For this, we propose a synthetic dataset of simple geometric shapes for controlled analysis, along with queries crafted to precisely test language cues. Furthermore, the dataset is modified to test causal reliance on visual evidence. Our results show that current VLMs combine genuine visual reasoning with shortcut strategies primarily rooted in language cues.