Inside a neutrino AI, useful physics can sit unused

IceCube reconstructs particle events from sparse patterns of Cherenkov light recorded by optical sensors in Antarctic ice. Machine-learning models can turn those pulse patterns into estimates of a neutrino’s direction. But their internal representations can contain much more information than a direction estimate reveals.

Raphaël Bonnet-Guerrini and colleagues investigated which recognizable properties PolarBERT stores internally, and whether its prediction components actually use them. Their study, arXiv:2608.26090v2, was revised on 6 November 2026. It is a preprint submitted to SciPost Physics; the supplied record does not establish journal acceptance or peer-reviewed publication.

Separating stored information from useful machinery

PolarBERT is an eight-block transformer pretrained on IceCube pulse sequences and subsequently fine-tuned to reconstruct direction. The researchers examined a 256-number internal summary of each event, which supplies the direction-prediction component, or “head,” with its only input.

They analyzed approximately one million simulated events from the public IceCube Kaggle sample. Separate subsets served different purposes: training an interpretability tool, checking its reconstruction quality, discovering candidate features, independently validating those features, and testing interventions.

The tool was a sparse autoencoder. It rewrites the model’s dense event summary using a larger dictionary of 1,024 possible features, with about 16 active per event on average. The hope is that some dictionary entries separate recognizable concepts that overlap in the original representation. The dictionary was trained without physics-concept labels, although interpreting and validating its features required additional measurements and controls.

A plausible-looking association was not enough. Candidates had to survive held-out tests and comparisons between events matched on potential confounding properties, including charge, pulse counts and, where applicable, simulated arrival direction. The team also repeated analyses across independently trained dictionaries.

A small, qualified atlas

The clearest final-layer feature responded to bright events dominated by non-auxiliary pulses. Another tracked auxiliary-dominated activity. Here, “auxiliary” refers to a detector readout flag, not a definitive label identifying noise. These pulses are more susceptible to uncorrelated noise, but the distinction matters when interpreting what the model has learned.

Brightness and auxiliary content were strongly correlated in this sample. The strongest feature also distinguished brighter from dimmer events within a clean-event subset, so the authors called it “bright-clean” rather than a pure quality indicator.

The atlas was incomplete. Most dictionary entries did not map to validated tested concepts. A strong detector-depth feature appeared in an earlier layer, not the final summary. Failure to isolate a concept in one location therefore did not show that the network lacked it.

Replication was also uneven. The primary bright-clean role passed all control schemes in three of four dictionary trainings. The individual auxiliary feature passed the full validation step only in the reference training. Broader response patterns were more stable than particular coordinates.

The decisive next step was intervention: removing features, transferring their activations between contrasting events, or changing their strength before passing the altered summary through a prediction head.

Removing the reference bright-clean feature changed direction-reconstruction error by only about 0.03 degrees, comparable to unrelated control interventions. Individual and grouped tests generally found no selective direction-head dependence on the validated atlas. This does not establish that direction reconstruction ignores all physical information—only that these identified readouts were largely unused under the tests performed.

A separate, dense coordinate strongly affected direction reconstruction, but failed concept-selectivity tests. Changing it upward or downward worsened reconstruction, so its influence could not simply be interpreted as the head reading “more brightness.”

Predicting which reconstructions to trust

The researchers then froze the existing model and trained a small uncertainty head to predict a logarithmic measure of its angular error. A linear regressor performed comparably, suggesting much of the relevant information was already readily accessible.

Unlike the direction head, this predictor depended on the clean-side features. Removing the bright-clean feature increased predicted error beyond matched controls, with the effect recurring across all four dictionaries. These interventions establish dependence within the model, not causation in the detector itself.

On the held-out intervention sample, retaining the 20% of events ranked best by the uncertainty head yielded a median angular error of about 3.2 degrees. Selecting the highest-charge 20% gave 20.2 degrees; the unselected sample’s median was 51.3 degrees. This is better selection, not improved direction estimates for individual events.

Important limits remain. The demonstration used simulation, not real detector data. Predicted errors ranked events well but were overconfident in an intermediate range and were not globally calibrated uncertainties. Interpretability did not establish a prediction-performance advantage over ordinary supervised readouts. Its contribution was to identify some physical information the new head actually uses—and show why storing information and using it are different claims.