On Fairness of Task Arithmetic: The Role of Task Vectors

Laura Gomezjurado Gonzalez ♯ \sharp, Hiroki Naganuma ♯ \sharp, Kotaro Yoshida ♯ \sharp, Yuji Naraki, Takafumi Horie, Ryotaro Shimizu · 2025

Read On Fairness of Task Arithmetic: The Role of Task Vectors by Laura Gomezjurado Gonzalez ♯ \sharp, Hiroki Naganuma ♯ \sharp, Kotaro Yoshida ♯ \sharp, Yuji Naraki, Takafumi Horie, Ryotaro Shimizu (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-2505-24262.txt.

Description

Hiroki Naganuma, Kotaro Yoshida, Laura Gomezjurado Gonzalez, Takafumi Horie, Yuji Naraki, and Ryotaro Shimizu study fairness in a shortcut for adapting AI models, in a 2025 paper. Task arithmetic edits a model by adding or subtracting small change vectors, skipping full retraining. The savings are real, but no one had checked whether the shortcut treats social groups fairly.

The team compare task arithmetic against full fine-tuning and a low-rank method on text and image sorting tasks, including hate speech detection. Standard fairness measures, like equal error rates across groups, grade each method. Their analysis shows the change vectors can be tuned to reach fairness on par with the heavier methods. Cheap editing that stays fair matters as these tools spread.

Catalog record

CREATORLaura Gomezjurado Gonzalez ♯ \sharp, Hiroki Naganuma ♯ \sharp, Kotaro Yoshida ♯ \sharp, Yuji Naraki, Takafumi Horie, Ryotaro Shimizu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE68 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-24262

Browse the full library · About the Tower of Babel