This paper points out that the main cause of the poor visual reasoning ability of the Vision-Language Model (VLM) is the failure of the binding problem between visual features and their targets. Existing VLMs mainly process visual features in parallel, and lack a spatially based sequential attention mechanism. To solve this problem, this paper presents a simple but effective method that adds low-level spatial structures (e.g., horizontal lines) to the visual input and uses text prompts to induce spatially aware sequential parsing. Experimental results show significant performance improvements in various visual reasoning tasks, including 25% visual search accuracy, 26.83% computation accuracy, 0.32 edit distance error reduction in scene description, and 9.5% performance improvement in spatial relation task. We confirm that purely language-based approaches (e.g., Chain-of-Thought prompting) are ineffective or even degrade performance, while visual modification is essential. The result of improving the binding problem with only single-query inference emphasizes the importance of visual input design. Low-level visual structuring represents a powerful and underexplored direction for improving constructive visual reasoning, suggesting that it could serve as a general strategy for improving VLM performance in space-based tasks.
