2026年04月15日 日本語

Can Foundation Models Grasp Space? Testing Spatial Beliefs with ‘Theory of Space’

Stanford AI Lab has unveiled ‘Theory of Space,’ a benchmark designed to test whether foundation models can construct, revise, and exploit spatial beliefs through active exploration. The study evaluated six state-of-the-art models, revealing a critical “exploration bottleneck,” a persistent “text-vision modality gap,” and severe “belief incoherence.”

Why is spatial understanding a challenge for AI? Theory of Space offers a novel standard for assessing an AI’s ability to grasp its position and relationships within physical space via interaction, moving beyond mere pattern recognition. This benchmark highlights the limitations in AI’s spatial comprehension and provides crucial insights for future research and development.