When a neutrino passes through ice, it leaves a faint flash of light. The IceCube detector captures it, and the PolarBERT model must guess where the particle came from. Previously, it was unclear what the model sees inside.

To find out, researchers applied a sparse autoencoder, a neural network that extracts a small number of meaningful features. The result was a dictionary of 1024 'shelves', each holding a distinct physical concept. For instance, one shelf is responsible for bright and clean events, another for the depth at which the flash occurred.

In contrast, the uncertainty head causally depends on the bright-clean features. By training it, scientists could select events so that directional accuracy improved from 20.2° to 3.2° at a 20% selection. Now it's clear: the model knows the physics but doesn't use it for the main task. This opens the door to new applications.