---
title: "Finding and Using Interpretable Latents in a Neutrino Foundation Model with Sparse Autoencoders"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.26090v1"
permalink: "https://feed.astralor.com/items/itm_c1468b290c9d4da93b4fbfe2"
published: "2026-08-26T17:53:00.000Z"
score: 78
channel: "ai"
tags: ["IceCube","Neutrino Foundation Model","Mechanistic Interpretability","Sparse Autoencoder","Particle Physics"]
---

# Finding and Using Interpretable Latents in a Neutrino Foundation Model with Sparse Autoencoders

## AI commentary

> What stands out most is not just another SAE application, but the demonstration that interpretability tools can guide physics task design—discovering underused physical concepts and exploiting them via an uncertainty head leads to an order-of-magnitude improvement in angular resolution. Researchers in AI interpretability and computational particle physics should read this carefully.

## At a glance

This work presents the first application of sparse-autoencoder-based mechanistic interpretability to particle physics. Studying a neutrino foundation model pretrained on IceCube data and fine-tuned for direction reconstruction, the authors identify a validated atlas of physical concepts via held-out tests, matched nuisance controls, and replication across independent dictionary trainings. Causal interventions show the direction head barely uses this atlas; motivated by this underused information, they train an uncertainty head on the same event-level representation to predict angular reconstruction error. At 20% selection efficiency, this interpretable estimator improves median angular resolution from 20.2° to 3.2°, suggesting mechanistic interpretability can reveal learned latent physics and guide downstream tasks.

## Key points

- First application of sparse-autoencoder mechanistic interpretability to a particle physics foundation model
- A validated atlas of physical concepts established via held-out tests, nuisance controls, and replication
- Direction head barely uses the atlas, while the uncertainty head causally depends on quality and brightness features
- Median angular resolution improves from 20.2° to 3.2° at 20% selection efficiency

## Full text

We present a first application of sparse-autoencoder-based mechanistic interpretability to particle physics. Studying a neutrino foundation model pretrained on IceCube data and fine-tuned for direction reconstruction, we identify a validated atlas of physical concepts in the model representation, using a strict validation protocol consisting of held-out tests, matched nuisance controls, and replication across independent dictionary trainings. Causal interventions show that the direction head barely draws on this atlas. Motivated by this underused information, we train an uncertainty head on the same event-level representation to predict the model's angular reconstruction error. Unlike the direction head, it depends causally on quality and brightness features from the atlas. At $20\%$ selection efficiency, this interpretable estimator improves the median angular resolution from $20.2^\circ$ to $3.2^\circ$. These results suggest that mechanistic interpretability can reveal learned latent physics encoded within a model's internal representation and help design downstream tasks that exploit it.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_c1468b290c9d4da93b4fbfe2