Equivalence of Context and Parameter Updates in Modern Transformer Blocks

Adrian Goldwaser [email protected], Michael Munn [email protected], Javier Gonzalvo [email protected], Benoit Dherin [email protected] · 2025

Read Equivalence of Context and Parameter Updates in Modern Transformer Blocks by Adrian Goldwaser [email protected], Michael Munn [email protected], Javier Gonzalvo [email protected], Benoit Dherin [email protected] (2025) 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-2511-17864.txt.

Description

Adrian Goldwaser, Michael Munn, Javier Gonzalvo, and Benoit Dherin show that the effect of a prompt's context on a modern transformer can be rewritten as a small change to its weights. Earlier work proved this for a basic transformer, using a rank one patch that depends on the input tokens.

For a Gemma style block they give an exact solution, patching both the main weight matrices and a normalization scale. A general framework then covers many layer types, including gated and mixture of experts blocks, under two controllability conditions. The result gives a simpler lens on how prompts reshape a model from inside.

Catalog record

CREATORAdrian Goldwaser [email protected], Michael Munn [email protected], Javier Gonzalvo [email protected], Benoit Dherin [email protected]
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE31 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-17864

Browse the full library · About the Tower of Babel