Towards Spectroscopy: Susceptibility Clusters in Language Models
Andrew GordonTimaeus [email protected], Garrett BakerTimaeus [email protected] George WangTimaeus [email protected] William SnellTimaeus [email protected] Stan van WingerdenTimaeus [email protected] Daniel MurfetTimaeus [email protected] · 2026
Read Towards Spectroscopy: Susceptibility Clusters in Language Models by Andrew GordonTimaeus [email protected], Garrett BakerTimaeus [email protected] George WangTimaeus [email protected] William SnellTimaeus [email protected] Stan van WingerdenTimaeus [email protected] Daniel MurfetTimaeus [email protected] (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-12703.txt.
Introduction
Andrew Gordon and five co authors at Timaeus borrow a method from physics to study language models. Spectroscopy infers internal structure by measuring how a system responds to perturbations. They apply that principle to neural networks. Their perturbation upweights one token in a given context, shifting the data distribution slightly.
The response is measured through susceptibilities, covariances between component level observables and the perturbation. These are computed over a localized Gibbs posterior, in the language of the work. A susceptibility map could show which parts of a model care about which tokens. That would give interpretability a tool with roots in statistical physics. The team frame this as a step toward a spectroscopy of models.
Catalog record
| CREATOR | Andrew GordonTimaeus [email protected], Garrett BakerTimaeus [email protected] George WangTimaeus [email protected] William SnellTimaeus [email protected] Stan van WingerdenTimaeus [email protected] Daniel MurfetTimaeus [email protected] |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 164 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-12703