GaussMark: A Practical Approach for Structural Watermarking of Language Models

Adam Block, Alexander Rakhlin, Ayush Sekhari · 2025

Read GaussMark: A Practical Approach for Structural Watermarking of Language Models by Adam Block, Alexander Rakhlin, Ayush Sekhari (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-2501-13941.txt.

Introduction

Adam Block and co-authors proposed a new way in 2025 to watermark language models by marking the model itself rather than its words. Most marking schemes tweak each token of output, which can slow writing, dull quality, or wash out under edits. A mark inside the model's weights could dodge those costs.

Their scheme, GaussMark, stirs a pinch of random noise into the weights. Only the owner, holding a secret key, can test for that exact noise pattern in new text's behavior. The team proves bounds on how sound and strong the test is. Wide trials showed the mark survives inserts, deletions, swaps, and even round-trip translation. Text quality stayed much the same. The authors call it a practical mark that costs nothing in writing speed.

Catalog record

CREATORAdam Block, Alexander Rakhlin, Ayush Sekhari
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE201 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-13941

Browse the full library · About the Tower of Babel