Vision language models (VLMs), the AI systems trained to understand both images and text, excel at analyzing single snapshots but struggle dramatically when they need to track how spatial relationships change over time. A new benchmark called MindTopo, developed by Microsoft Research, reveals this gap by testing whether multimodal AI can understand and preserve topological properties like connectivity, enclosure, and knottedness as objects move and scenes evolve.

What Is Topology, and Why Should You Care About AI Understanding It?

Topology is a form of spatial reasoning that focuses on structural relationships that remain true even when objects bend, stretch, or deform. Unlike Euclidean geometry, which measures exact distances and angles, topology asks simpler but more fundamental questions: Are two rooms still connected after a wall is added? Is an animal inside a fence or outside? Is a rope truly knotted or just tangled? These concepts form a foundational layer of how humans understand space, yet they have been largely absent from how AI systems are evaluated.

The implications are significant. Robots, accessibility tools, and interactive AI assistants must understand not only where objects are located but also what remains connected, enclosed, ordered, or knotted as actions unfold. Without this capability, AI systems can make decisions that violate physical constraints or fail to accomplish tasks that require maintaining spatial relationships over multiple steps.

How Does MindTopo Test AI Spatial Reasoning?

The benchmark organizes its evaluation around five core topological categories, each tested at two cognitive levels. The categories include continuity, which asks whether a path or object remains unbroken; separation, which tests whether nearby elements form one structure or distinct parts; order, which tracks how elements are arranged along a path or transformation; enclosure, which tests whether a boundary creates an inside and an outside; and knots, which determines whether ropes are truly knotted or linked rather than merely tangled in appearance.

For each category, researchers test both reasoning and planning. In reasoning tasks, a model examines one or more rendered scenes and answers a question about their topological structure. In planning tasks, the model interacts with a simulated environment and must select actions that create, preserve, or remove a particular relationship, such as rotating pipe segments, drawing a separating path, rearranging blocks, or untangling ropes. The environments enforce legal actions, so a model cannot cheat by passing one rope strand through another.

What Did the Research Reveal About Current AI Limitations?

The findings expose a striking performance gap. Across a broad set of proprietary and open-weight models, performance was consistently stronger on static reasoning than on interactive planning, and both remained well below human performance. The contrast was especially clear when success depended on preserving a relationship across many actions.

The error patterns help pinpoint where models fail. Static mistakes usually began with perception, such as missing a wall, opening, or crossing. Planning mistakes appeared after the scene had been understood. Models followed a locally plausible move without tracking its later consequences, lost the task over multiple turns, or proposed an action that violated the environment's dynamics.

Researchers also tested whether image and video generation could help models maintain an understanding of topological relationships. Image generation sometimes helped when the relevant relation was visible in a single frame, but it remained unreliable across a sequence of crossings or moves. Video rollouts frequently altered topology or violated task dynamics. Visual simulation appeared useful only to the extent that it preserved structural constraints over time.

How Can AI Systems Improve Their Spatial Reasoning?

• Explicit Topological State: Models may need to carry an explicit representation of topological properties, allowing them to track which relationships must be preserved as scenes change and actions unfold.
• World Models with Structural Constraints: Developing world models whose predictions preserve topology by construction could ensure that AI systems never propose actions that violate spatial relationships or physical laws.
• Planning-Focused Training: Since failures emerge during planning rather than perception, training approaches that emphasize multi-step reasoning and consequence tracking may help models maintain spatial understanding over longer action sequences.

MindTopo is intended as a controlled diagnostic for this gap. By using scenes generated from controlled simulators, researchers can provide exact ground truth and adjustable difficulty, making it possible to separate two failure modes that otherwise look alike: a model that fails because a scene is visually complex, and a model that fails because it cannot maintain the underlying relationship as objects move.

What Are the Real-World Implications for AI Applications?

The findings have direct implications for robotics, accessibility tools, and interactive assistants. A robot that cannot reliably track whether a door remains connected to its frame after a sequence of movements might fail to navigate a building. An accessibility tool that loses track of spatial relationships might give incorrect instructions. An interactive assistant that cannot maintain topological understanding might propose physically impossible solutions to problems.

The research highlights an important opportunity to advance AI systems for environments where understanding spatial relationships is essential for reliable decision-making. As multimodal AI becomes increasingly integrated into physical-world applications, the ability to maintain topological intuition across time and action sequences will become a critical capability, not a nice-to-have feature.