Beyond 3D VQAs: Injecting 3D Spatial Priors into Vision-Language Models for Enhanced Geometric Reasoning

Researchers introduce GASP, a framework that enhances Vision-Language Models' 3D spatial reasoning by injecting geometric priors directly into transformer layers rather than relying on 3D VQA datasets. The approach uses contrastive learning on point correspondences and depth consistency supervision, achieving 70%+ correspondence accuracy and 18-29% improvements on spatial benchmarks without any 3D VQA training data.

GASP addresses a fundamental limitation in how Vision-Language Models approach 3D spatial understanding. Rather than fine-tuning on task-specific datasets—an approach prone to overfitting and poor generalization—the framework teaches models to learn geometric principles from first principles. This shift in training methodology reflects a broader maturation in AI research, moving away from narrow benchmarks toward more robust foundational capabilities.

The technical innovation centers on embedding geometric awareness throughout the model's architecture. By applying supervision across all transformer layers rather than just the output, GASP forces the model to develop consistent spatial representations internally. The diagnostic finding that standard VLMs achieve below 5% correspondence accuracy reveals a critical gap in current architectures that this approach directly addresses.

For AI developers and researchers, this represents a scalable pathway to improve spatial reasoning without architectural redesigns or specialized 3D encoders. The ability to achieve substantial downstream improvements (18-29% gains) without 3D VQA training data demonstrates that geometric priors learned from video scenes transfer effectively to diverse spatial reasoning tasks. This efficiency appeals to organizations with limited computational budgets.

The implications extend to applications requiring reliable 3D understanding—robotics, autonomous systems, and spatial planning tools. As VLMs increasingly power perception systems in real-world applications, fundamental geometric competence becomes critical. Future work likely explores scaling this approach across larger models and more diverse geometric scenarios, establishing geometric literacy as a baseline expectation in production VLMs.

• →GASP injects geometric priors into VLMs via contrastive learning on point correspondences, improving internal correspondence matching from <5% to >70%
• →The framework achieves 18-29% improvements on spatial benchmarks without training on any 3D VQA datasets, demonstrating strong generalization
• →Standard Vision-Language Models show critically low internal spatial understanding, revealing a fundamental architectural limitation addressed by deep layer supervision
• →Learning from geometric principles rather than task-specific datasets reduces overfitting and enables more transferable spatial reasoning capabilities
• →The approach scales efficiently without requiring specialized 3D encoders, making it practical for integration into existing VLM architectures