MLLMs Face 'Cartesian Illusion' in Spatial Reasoning Tasks

A new paper on arXiv (2605.18194) exposes a fundamental limitation in Multi-Modal Large Language Models (MLLMs): their spatial intelligence is hampered by a 'Cartesian Illusion,' meaning they rely on text-based probability distributions rather than grounded 3D topological understanding. This deficiency becomes critical in multi-agent environments requiring second-order Theory of Mind (ToM)—where an agent must infer another agent's beliefs based on its physical orientation and sensory limits. The authors probe these limits with a novel audio-visual task: Agent A predicts Agent B's estimate of A's location. To address this, they propose an Epistemic Sensory Bottleneck module that avoids rigid coordinate transformations, instead using an Anchor mechanism. The research highlights that current MLLMs lack embodied spatial reasoning, which is essential for tasks like navigation and human-robot interaction.