Visual Structures Help Visual Reasoning: Addressing the Binding Problem in VLMs

NeurIPS 2025
Amirmohammad Izadi∗1, Mohammad Ali Banayeeanzade∗1, Fatemeh Askari1, Ali Rahimiakbar1, Mohammad Mahdi Vahedi1, Hosein Hasani1, Mahdieh Soleymani Baghshah1
∗Equal contribution.
1Department of Computer Engineering, Sharif University of Technology
NeurIPS NeurIPS page arXiv
Main Figure
Viser Visual Structures for Better Visual Reasoning — we add light-weight row-wise visual structure and a short prompt to guide sequential scanning, improving object counting by addressing the binding problem.
strawberries
1
2
3
4
How many strawberries are there in this image?
Total: 10×
Viser: Scan the image sequentially based on the horizontal lines.
Row 1: 2 strawberries
Row 2: 4 strawberries
Row 3: 3 strawberries
Row 4: 2 strawberries
Total: 11✓

Abstract

Despite progress in Vision-Language Models (VLMs), their capacity for visual reasoning is often limited by the binding problem: the failure to reliably associate perceptual features with their correct visual referents. This limitation underlies persistent errors in tasks such as counting, visual search, scene description, and spatial relationship understanding. A key factor is that current VLMs process visual features largely in parallel, lacking mechanisms for spatially grounded, serial attention.

This paper introduces VISER (Visual Input Structure for Enhanced Reasoning), a simple yet effective intervention: augmenting visual inputs with low-level spatial structures and pairing this with a textual prompt that encourages sequential, spatially aware parsing. We empirically demonstrate substantial performance improvements across core visual reasoning tasks. Specifically, VISER improves GPT-4o visual search accuracy by 25.00%, increases counting accuracy by 26.83%, reduces edit distance error in scene description by 0.32, and enhances performance on spatial relationship tasks by 9.50% on a 2D synthetic dataset.

Furthermore, we find that the visual modification is essential for these gains; purely textual strategies, including Chain-of-Thought prompting, are insufficient and can even degrade performance. VISER enhances binding using only a single-query inference, underscoring the importance of visual input design over purely linguistically based approaches. These findings suggest that low-level visual structuring is a powerful and underexplored direction for improving compositional visual reasoning and could serve as a general strategy for enhancing VLM performance on spatially grounded tasks.

Results

Table 1: Counting accuracy across models and datasets
Table 2: Visual search harmonic mean results
Table 3–4: Scene description and spatial relationship results

BibTeX

@article{izadi2025visual,
title = {Visual Structures Help Visual Reasoning: Addressing the Binding Problem in VLMs},
author = {Izadi, Amirmohammad and Banayeeanzade, Mohammad Ali and Askari, Fatemeh and Rahimiakbar, Ali and Vahedi, Mohammad Mahdi and Hasani, Hosein and Soleymani Baghshah, Mahdieh},
journal = {arXiv preprint arXiv:2506.22146},
year = {2025}
}