Stanford University researchers have introduced "Theory of Space," a novel benchmark designed to evaluate whether artificial intelligence can build and maintain mental maps of physical environments. Inspired by the cognitive science concept of "Theory of Mind," this framework shifts the focus from passive observation to active exploration. Instead of simply processing a static log of images, AI agents must autonomously decide where to move and what to look at to understand their surroundings.

The study tested leading models, including GPT-5.2 and Gemini 3 Pro, across three core dimensions: constructing a spatial belief, exploiting that belief to solve tasks, and revising it when the environment changes. The results highlight a significant "modality gap," where models perform much better when receiving text-based descriptions compared to visual data. Even the most advanced systems struggle with "belief inertia," which is the tendency to cling to outdated spatial information even after seeing that an object has been moved.

These findings suggest that while today’s foundation models are excellent at processing provided data, they lack the efficiency and stability required for real-world robotics. Current agents often take significantly more steps than human-scripted versions to map a room, frequently leaving areas partially explored. By externalizing internal cognitive maps, the Theory of Space benchmark provides a clear roadmap for developing AI that can truly understand the three-dimensional world it inhabits.

Stanford University researchers have introduced "Theory of Space," a novel benchmark designed to evaluate whether artificial intelligence can build and maintain mental maps of physical environments. Inspired by the cognitive science concept of "Theory of Mind," this framework shifts the focus from passive observation to active exploration. Instead of simply processing a static log of images, AI agents must autonomously decide where to move and what to look at to understand their surroundings.

The study tested leading models, including GPT-5.2 and Gemini 3 Pro, across three core dimensions: constructing a spatial belief, exploiting that belief to solve tasks, and revising it when the environment changes. The results highlight a significant "modality gap," where models perform much better when receiving text-based descriptions compared to visual data. Even the most advanced systems struggle with "belief inertia," which is the tendency to cling to outdated spatial information even after seeing that an object has been moved.

These findings suggest that while today’s foundation models are excellent at processing provided data, they lack the efficiency and stability required for real-world robotics. Current agents often take significantly more steps than human-scripted versions to map a room, frequently leaving areas partially explored. By externalizing internal cognitive maps, the Theory of Space benchmark provides a clear roadmap for developing AI that can truly understand the three-dimensional world it inhabits.