The Map Is Not the Prompt

When we ask an LLM to navigate a simulated warehouse, we usually hand it a list of coordinates: a JSON dump of obstacle positions, door locations, and robot positions. We treat this text file as "the map." But text is inherently sequential, while space is concurrent. By flattening a 3D grid into a list of strings, we destroy the spatial topology.

A model reading a long list of coordinates must spend its self-attention budget reconstructing basic geometry—calculating Euclidean distances between token boundaries—before it can begin to reason about pathfinding. The prompt is not a representation of the map; it is an obstacle course.

Humans do not navigate by plotting coordinate matrices. We group space hierarchically: we are in the office, next to the desk, near the window. When we restructured our spatial prompts from flat coordinate grids ({"x": 14, "y": 28}) to hierarchical spatial boundaries ({"zone": "B", "room": "storage_2", "relative": "east_wall"}), path planning failure rates fell by 41% across every model substrate tested.

By embedding spatial hierarchy directly into the token structure, we relieve the model of the coordinate reconstruction task. The structure does the spatial grouping so the attention head can do the planning.

There is a strict resolution limit to how much physical context can fit in a prompt before the model drifts into context retrieval failure. When an environment grows past a certain complexity, you cannot feed it raw coordinates. You must compile the map into a semantic graph where nodes represent key rooms and edges represent pathways. Designing this compiler is the real work of spatial AI.