Spherical Boltzmann machines: a solvable theory of learning and generation in energy-based models
Thomas Tulinski, Simona Cocco, Rémi Monasson, Jorge Fernandez-de-Cossio-Diaz · 2026
Read Spherical Boltzmann machines: a solvable theory of learning and generation in energy-based models by Thomas Tulinski, Simona Cocco, Rémi Monasson, Jorge Fernandez-de-Cossio-Diaz (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-2605-09031.txt.
Introduction
Thomas Tulinski works with three co-authors, Simona Cocco, Rémi Monasson, and Jorge Fernandez-De-Cossio-Diaz, on a solvable energy-based model of learning and generation. Their 2026 study examines the spherical Boltzmann machine, a flexible model inspired by statistical physics whose training behavior is usually hard to pin down. Working in the high-dimensional limit, they combine random matrix tools with dynamical mean-field methods.
The team solves exact equations for how training unfolds and computes a quantity that summarizes global traits of the trained model. They uncover chains of phase transitions during training, tied to the coupling matrix aligning step by step with patterns in the data. In a teacher-student setup they link those transitions to sampling effects, including double descent and biases from unfinished training. Tests suggest the same effects appear in standard generative systems, so the solvable model offers a window onto broader practice.
Catalog record
| CREATOR | Thomas Tulinski, Simona Cocco, Rémi Monasson, Jorge Fernandez-de-Cossio-Diaz |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 177 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.09031