Vision-language models (VLMs) have achieved strong performance on a wide range of visual understanding tasks. Yet it remains unclear whether these models genuinely learn spatial relationships or primarily rely on correlations between visual patterns and linguistic labels.

We introduce SPAR, a benchmark designed to distinguish relational reasoning from surface-level visual cues. SPAR contains 48K image-question pairs covering six spatial relations: left/right, above/below, inside/outside, near/far, overlapping, and disconnected.

Across three model scales, we find that spatial reasoning performance improves substantially with scale, with the largest gains appearing on compositional examples. These results suggest that sufficiently large vision-language models can develop an emergent ability to reason about spatial relationships.

Vision-language models (VLMs) have made substantial progress in visual question answering and multimodal understanding. However, strong benchmark performance does not necessarily imply that models have acquired the underlying concepts required for reasoning about spatial relationships.

Recent work suggests that larger models can develop capabilities that are difficult to observe in smaller models.¹ Whether such improvements reflect genuine relational reasoning, however, remains unclear.

We argue that existing benchmarks often make it difficult to separate spatial reasoning from recognition of familiar visual-language patterns. For example, a model may correctly answer that a red object is left of a blue object without constructing an explicit representation of their relative positions.

To address this problem, we introduce SPAR, a benchmark designed around compositional spatial relations.

SPAR contains 48K image-question pairs generated from structured scenes. Each scene contains between two and eight objects with controlled attributes, positions, and relations.

We divide the benchmark into two categories:

1. Single-relation examples, which require identifying one spatial relationship between two objects.
2. Compositional examples, which require combining two or more relations to answer the question.

For example, given that object A is left of object B and object B is above object C, the model may be asked to determine the relative position of A and C.

To reduce reliance on superficial cues, object identities, colors, and spatial configurations are randomized across examples.

We evaluate three model scales, referred to as Small, Medium, and Large. All models are trained using the same optimization procedure and evaluated on the same held-out test set.

3.1 Results

Scaling consistently improves performance across all six spatial relations. However, the improvement is substantially larger on compositional examples than on single-relation examples.²

The performance gap between single-relation and compositional examples decreases as model scale increases. This pattern suggests that larger models are better able to combine multiple spatial relations.

3.2 Ablation Study

We next remove compositional examples from the training set while keeping the total number of training examples constant.

Interestingly, models trained without compositional examples still exhibit substantial improvements on the compositional test set as model scale increases.

These results indicate that exposure to compositional examples is not strictly necessary for the emergence of spatial reasoning.

Our experiments show that spatial reasoning performance improves with model scale, particularly on tasks requiring the composition of multiple relations.

We interpret this result as evidence that sufficiently large VLMs can acquire an internal representation of spatial relationships without requiring explicit supervision for every possible composition.

An alternative explanation is that larger models become better at exploiting statistical regularities in the benchmark. Although SPAR reduces several common surface-level cues, we cannot completely rule out this possibility.

Future work should evaluate whether the observed improvements transfer to natural images and spatial configurations that differ substantially from those encountered during training.

We introduced SPAR, a benchmark for studying spatial reasoning in vision-language models. Our results demonstrate that scaling substantially improves performance on compositional spatial tasks.

Taken together, the findings suggest that spatial reasoning may emerge as a natural consequence of increasing model capacity, even in the absence of explicit relational supervision.