GeoEx feature-plane extraction

Train a probe constrained to an affine transform with an L1-sparsified scaling vector, so that it must predict a continuous property from a low-dimensional plane of the hidden space. The plane can then be read, or pushed on by moving every latent vector's projection onto a chosen point and generating from the edited state. Moving from reading the plane to pushing on it is what turns a correlational claim into an interventional one.

Findings

• TM-011Spatial coherence of latent feature planes increases with encoder depth [primary]
• TM-012Longitude and time-of-year planes become circular in deeper encoder blocks [primary]
• TM-013Intervening on the altitude plane raises generated terrain by almost 500 metres [primary]
• TM-014Per-edge latent distance along shortest paths spikes at physical barriers [supporting]