SpatialScore: Towards Comprehensive Evaluation for Spatial Intelligence

Researchers introduce SpatialScore, a comprehensive benchmark with 5K samples across 30 tasks to evaluate multimodal language models' spatial reasoning capabilities. The work includes SpatialCorpus, a 331K-sample training dataset, and SpatialAgent, a multi-agent system with 12 specialized tools, demonstrating significant improvements in spatial intelligence without additional model training.

The development of SpatialScore addresses a critical gap in AI evaluation frameworks. Existing benchmarks for spatial understanding in multimodal large language models remain fragmented and narrow in scope, leaving researchers without standardized metrics to assess how well these systems comprehend spatial relationships, geometric reasoning, and three-dimensional concepts. This benchmark represents a shift toward more rigorous, comprehensive evaluation standards that mirror real-world spatial reasoning demands.

Spatial intelligence constitutes a foundational capability for autonomous systems, robotics, computer vision applications, and embodied AI. As MLLMs become increasingly central to commercial AI products, their ability to process spatial information accurately determines viability for navigation systems, architectural visualization, manufacturing automation, and other spatial-dependent industries. The evaluation of 49 representative models reveals persistent capability gaps, suggesting current architectures struggle with spatial abstractions that humans process intuitively.

The research demonstrates two complementary advancement pathways. The data-driven approach via SpatialCorpus shows measurable performance improvements through fine-tuning, exemplified by Qwen3-VL's enhanced spatial reasoning. Alternatively, SpatialAgent's training-free paradigm using specialized perception tools provides immediate performance gains without computational overhead, making it accessible to researchers with limited resources. This dual-route methodology acknowledges that capability gains don't require full model retraining.

For the AI industry, these resources establish benchmarks for evaluating spatial reasoning capabilities in next-generation models. Organizations developing autonomous systems, robotics platforms, and spatial computing applications gain standardized evaluation tools. The public release of data, code, and models accelerates research democratization, enabling broader participation in advancing spatial AI. The framework's effectiveness suggests spatial reasoning remains a distinct capability requiring targeted improvement strategies rather than general scaling.

• →SpatialScore introduces the most comprehensive spatial intelligence benchmark with 5K manually verified samples across 30 distinct tasks for evaluating MLLMs.
• →Evaluation of 49 models reveals substantial gaps between current spatial reasoning capabilities and human-level performance in multimodal understanding.
• →SpatialCorpus's 331K training samples significantly improve model performance, with Qwen3-VL demonstrating measurable enhancements through fine-tuning.
• →SpatialAgent's training-free multi-agent system with 12 specialized tools achieves substantial performance gains without requiring model retraining.
• →Public release of benchmark, corpus, and agent framework democratizes spatial AI research and establishes standardized evaluation metrics for the field.