In the Antarctic ice, a ghost particle leaves behind only a faint flash of blue light. The IceCube detector captures these flashes, and from them we need to infer where the particle came from. The PolarBERT model has learned to do this, but what exactly it sees inside has remained a mystery.

To peek inside, the researchers applied a sparse autoencoder, a neural network that extracts a small number of meaningful features from the model's activations. The result was a dictionary of 1024 latent directions. Each direction is like a shelf in a library, holding a physical concept.

But here is the paradox: the direction head, which should use this library, barely consults it. Removing or overwriting these features changes the predicted direction by less than 0.1°, as if a GPS navigator ignored a detailed map and followed only a rough line.

In contrast, the uncertainty head is causally dependent on the bright-clean features. Removing one latent changes the predicted error by a factor of 1.36. By training this head, the researchers could select events so that the median angular resolution improved from 20.2° to 3.2° at 20% selection efficiency.

The physics library turned out to be written, but not for the task it was built for. This opens the door to new applications: if a model stores knowledge it does not use, we can design a task that does use it. Then a flash in the Antarctic ice will tell us about its source much more precisely.