MindTopo tests the topological intuition of foundation models, visual generation deceives planning and falls behind human performance

The research community typically measures the capabilities of large visual and language models by relying on benchmark tests that focus on abstract geometric dimensions such as distances, angles, and visible shapes, yet true spatial perception, as grounded by cognitive science, is fundamentally based on topological relationships that remain invariant under continuous physical distortions and transformations. In this context, a research team introduces the MindTopo metric to examine topological intuition in foundation models across five primary properties derived from cognitive science and visual mathematical topology: connectivity, separation, ordering, enclosure, and understanding of nodes and their interconnections.

The new test partitions model evaluation into two distinct cognitive levels that delineate the boundary between theoretical understanding and practical execution. The first level is inference, where the model is asked to detect existing topological relationships in a scene or infer how they change and evolve; the second level is planning, in which the foundation model acts as a closed-loop software agent that determines its policy and motor actions within the environment. The metric comprises 11,030 test cases distributed across 13 procedurally generated task patterns with precisely controllable difficulty levels, and 14 multimodal models were subjected to the test, alongside an examination of architectures augmented with image generation and three video-generation models in closed-loop planning environments.

The results revealed a deep gap separating the models' theoretical perception from their inability to apply it in practice, whereAll fourteen models showed a clear superiority in inference tasks compared with their modest capability for executive planningwhile the performance of the best model under test remained far behind the observed human performance level on the same tasks. Even when evaluating the Qwen3-VL-2B-Instruct model and subjecting it to fine-tuning techniques under supervised and reinforcement learning, all gains went to mental inference, and the model's responses in motor planning trajectories remained unable to keep pace with the required improvement.

The most concerning observation appeared in video-based visual generation experiments, where models produce scenes that retain positional visual cues and reach endpoints that initially appear logical and convincing. However, rigorous scrutiny of successive motion trajectories proved thatthe generated scenes fail to follow the dynamics of the real environment and are unable to preserve correct topological properties when transitioning between frames, meaning that the model creates a deceptively successful scene while the true physical links dissipate beneath a superficial visual simulation.

This gap necessitates an immediate engineering review of software teams and companies developing robotic applications and automated logistics services in warehouses in Riyadh and Dubai or industrial sorting facilities in Cairo. Relying on generative models and vision-language models to make independent motor planning decisions entails a costly risk, as the experiment shows that a model's ability to answer scene questions never guarantees its ability to move a robotic arm or avoid wire entanglement and motion paths. The safest engineering choice here is to confine visual language models to inference and monitoring tasks, keeping planning and motor execution restricted to engines that enforce strict physical and mathematical verification of motion continuity in each frame before delegating field actions to the machine.

Our reading of these data shows that moving from experimental model environments to real-world field automation in the region requires deep scrutiny of how machines comprehend space and physical connectivity. Topological intuition is not a theoretical luxury but the governing condition for preventing operational accidents and routing errors in factories and warehouses whose rapidly accelerating digital economies depend on precise motion and overall system safety.