Spatial Intelligence in Vision-Language Models: A Comprehensive Survey

Spatial IntelligenceCognitive EnhancementAI Urban Planning

Vision-language models have achieved impressive progress, yet they still struggle with spatial intelligence.

TL;DR
• A comprehensive survey of spatial intelligence in vision-language models.
• Recent advances in spatially enhanced VLMs have produced a scattered literature with inconsistent terminology, methods, and evaluation practices.
• The survey provides a unified overview of core concepts, analysis of spatial failures, and organization of existing solutions into a clear framework.

Summary
Vision‑language models (VLMs) have advanced rapidly, yet spatial reasoning—understanding object locations, relationships, and viewpoint changes—remains a critical gap for embodied AI, autonomous driving, and consistent generation. This survey consolidates recent progress, identifies failure modes, and proposes a structured roadmap.
• Prompting‑based methods improve spatial grounding by incorporating explicit spatial cues.
• Model‑centric enhancements, such as 3D backbones and multi‑view fusion, yield higher accuracy on perception tasks.
• Data‑centric strategies, including curated spatial datasets and benchmark diversification, expose performance gaps and bias.
• Evaluation frameworks that separate perception, understanding, and extrapolation levels clarify where each technique excels.
• Open‑source code and a curated paper repository enable reproducibility and community collaboration.

The survey categorizes solutions into five pillars: prompting, model architecture, explicit 2D cues, 3D enrichment, and data‑driven approaches. Empirical analysis across 37 models and nine benchmarks reveals that multi‑view and 3D‑aware models consistently outperform single‑image baselines on perception tasks, while prompting techniques best address relational reasoning.

Benchmark analysis uncovers significant performance disparities across datasets, highlighting design biases in current spatial understanding tests. The authors recommend expanding benchmark coverage to include dynamic reasoning and perspective‑taking, ensuring that future models generalize beyond static scenes.

A practical roadmap is presented, outlining actionable steps for researchers: adopt multi‑view training pipelines, integrate scene‑graph prompts, and contribute to open datasets. The accompanying evaluation code facilitates rapid iteration and comparative studies, fostering a collaborative ecosystem.

By unifying terminology, evaluation practices, and solution frameworks, this work equips the community to accelerate spatial intelligence in VLMs, ultimately enabling safer autonomous systems and more reliable multimodal AI applications.

By consolidating evidence and outlining open challenges, this survey offers a practical roadmap for building more spatially capable VLMs.