Multi-Concept Editing Using Task Arithmetic

Niv Cohen, Nicky Kriplani, Benjamin Feuer, Yuval Lemberg, Chinmay Hegde

Keywords: Task Vectors

TL;DR: We study how specific task vectors affect a foundation model when applied together

Abstract:

Model owners often wish to introduce new capabilities into their trained models or remove undesired ones. Task Vectors (TVs) present a promising new approach to editing models after training, allowing simple and controllable addition of new capabilities to the model and the removal of undesired ones. But what happens when the model owner wants to change multiple capabilities?

In this work, we study the interactions of task vectors in a multi-edit setting for image classifiers and diffusion models. We start by quantifying the overall model degradation induced by applying many specific TVs simultaneously.
We show that the overall model performance degrades rapidly as the quantity of TV edits increases.
Finally, we explore different ways to mitigate this degradation and present an adaptive method to select the most relevant TVs to apply to a diffusion model during inference. Our technique achieves a 94.6% ROC AUC in identifying the correct TV, enabling the effective integration of multiple TV edits while significantly mitigating quality degradation.

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2724