The Fragment Map

Building a cognitive map seems to require traversing the space. Walk through a building, drive through a city, and the path experience compiles into a spatial representation. The assumption is sequential: the map forms from continuous trajectories, each step connecting to the next, gradually filling in the territory.

He, Daigle, and Bashivan show that this assumption is wrong. Their Episodic Spatial World Model constructs accurate spatial representations from sparse, disjoint episodes — disconnected snapshots of transitions between locations, never assembled into a continuous path. The model predicts previously unobserved transitions from minimal data, and its internal representations align geometrically with the actual environment structure.

The mechanism draws on how hippocampal episodic memory works in animals. Each episode records a local transition — from here to there under these conditions. The episodes are stored independently, not stitched into a narrative. The spatial structure emerges when the system infers relationships between episodes that share locations or obey consistent spatial rules. You don’t need to have walked from A to C. If you’ve walked from A to B and separately from B to C, the model infers the A-to-C structure.

This enables rapid adaptation. When the environment changes, only the affected episodes need updating — no global recomputation of the map. The system achieves near-optimal exploration and navigation between arbitrary locations without additional training, all from fragmentary input.

The structural point is about what maps require. Not coverage — completeness of traversal. Not continuity — unbroken sequential experience. Maps require consistency — episodes that obey the same spatial rules. Given consistency, fragments suffice. The gaps fill themselves because space is regular. What looks like incomplete knowledge is actually sufficient knowledge of a regular domain.