This paper proposes VISER (Visual Input Structure for Enhanced Reasoning) to address the limitations of visual language models (VLMs) in their visual reasoning capabilities. VLMs struggle to reliably connect perceptual features with visual referents, leading to errors in tasks such as computation, visual search, scene description, and spatial relationship understanding. VISER is a simple yet effective method for augmenting visual input with low-level spatial structure and adding text prompts that guide sequential and spatially aware parsing. Experimental results demonstrate that VISER significantly improves the performance of various visual reasoning tasks. Specifically, it improves visual search accuracy by 25.00% and computation accuracy by 26.83% on GPT-4o, reduces edit distance error in scene description by 0.32, and improves spatial relationship performance on a 2D synthetic dataset by 9.50%. This highlights the importance of visual input design over purely linguistic approaches and suggests that low-level visual structuring represents a powerful and unexplored direction for enhancing constructive visual reasoning.
