The Geometry of Transformer Weights – A Canonical Basis for LLMs

todot_ge1 pts0 comments

The Hidden Geometry of Transformer Weights: A Canonical Basis for Interpreting Transformer Language Models | Zenodo

Skip to main

You are using an outdated browser. Please upgrade your browser to improve your experience.

Published August 14, 2026

| Version v2

Preprint

Open

The Hidden Geometry of Transformer Weights: A Canonical Basis for Interpreting Transformer Language Models

Authors/Creators

Gernone, Gianluca

Description

Version 2 of the paper "The Hidden Geometry of Transformer Weights".

Rotating a Transformer language model into a coordinate system aligned<br>with its own weight matrices — the canonical basis — reveals a hidden<br>internal structure. Version 1 established the descriptive phenomena.<br>Version 2 establishes that they are functional and general:

- Causal evidence: zeroing a single canonical axis (0.11% of the model)<br>collapses MMLU from 47.50% to 21.25% and destroys output coherence;<br>five control axes show no effect.<br>- Six per-layer spectral indices (cohesive, torsional, informational,<br>dimensional, rhythmic, vorticity) with effective dimensionality 4.77/6,<br>exposing a 41x isotropic collapse between weight and activation spectra.<br>- Lossless realignment verified on RMSNorm and LayerNorm architectures<br>(Qwen 2.5 0.5B, SmolLM2 1.7B, Pythia 1.4B), plus native cross-layer<br>alignment measurements across eight architectures and a MoE model.<br>- Corrected cross-layer alignment numbers vs. version 1.

Every measurement is reproducible: scripts and pre-computed data<br>accompany the paper (https://github.com/todotge/canonical-basis),<br>including an interactive per-axis control chat<br>(demo video: https://youtu.be/WOJwkjj9VT0).

Files

A_Canonical_Basis_for_Interpreting_Transformer_Language_Models.pdf

Files<br>(463.7 kB)

Name<br>Size

Download all

A_Canonical_Basis_for_Interpreting_Transformer_Language_Models.pdf

md5:3c7758bd1010d5eff37bc3110a6388c1

463.7 kB

Preview

Download

Additional details

Related works

Is supplemented by

Working paper:

https://github.com/todotge/canonical-basis

(Other)

Dates

Updated

2026-08-14

Software

Repository URL

https://github.com/todotge/canonical-basis

Programming language

Python

156

Views

93

Downloads

Show more details

All versions<br>This version

Views

Total views

156

Downloads

Total downloads

93

Data volume

Total data volume

66.3 MB<br>0 Bytes

More info on how stats are collected....

Versions

External resources

Indexed in

OpenAIRE

Communities

Keywords and subjects

Keywords

transformer interpretability

canonical basis

singular value decomposition

model compression

sparse autoencoders

homeostasis

neural network geometry

layer normalization

mechanistic interpretability

ablation

MoE

Details

DOI

DOI Badge

DOI

10.5281/zenodo.21935673

Markdown

[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.21935673.svg)](https://doi.org/10.5281/zenodo.21935673)

reStructuredText

.. image:: https://zenodo.org/badge/DOI/10.5281/zenodo.21935673.svg<br>:target: https://doi.org/10.5281/zenodo.21935673

HTML

Image URL

https://zenodo.org/badge/DOI/10.5281/zenodo.21935673.svg

Target URL

https://doi.org/10.5281/zenodo.21935673

Resource type<br>Preprint

Publisher<br>Zenodo

Languages

English

Rights

License

Creative Commons Attribution 4.0 International

The Creative Commons Attribution license allows re-distribution and re-use of a licensed work on the condition that the creator is appropriately credited.

Read more

Copyright

Copyright (C) 2026 Gianluca Gernone

Citation

Export

Technical metadata

Created

August 14, 2026

Modified

August 14, 2026

Jump up

This site uses cookies. Find out more on how we use cookies

Accept all cookies<br>Accept only essential cookies

zenodo https canonical transformer basis version

Related Articles