Edit at your own risk: evaluating the robustness of edited models to distribution shifts

TL;DR

The paper examines how post-hoc model edits impact robustness to distribution shifts, addressing a gap where validation accuracy dominates evaluation. It develops a robustness-aware editing framework across two task families and introduces 1-layer interpolation (1-LI) to balance original and edited-task performance. Empirical results show that edits generally degrade robustness, with the degradation depending on editing method and layer; importantly, 1-LI often mitigates this degradation and reveals a monotone linear interpolation phenomenon (MLI) extending to single-layer edits. These findings offer practical guidance for choosing editing strategies that preserve robustness in real-world deployments facing distribution shifts.

Abstract

The current trend toward ever-larger models makes standard retraining procedures an ever-more expensive burden. For this reason, there is growing interest in model editing, which enables computationally inexpensive, interpretable, post-hoc model modifications. While many model editing techniques are promising, research on the properties of edited models is largely limited to evaluation of validation accuracy. The robustness of edited models is an important and yet mostly unexplored topic. In this paper, we employ recently developed techniques from the field of deep learning robustness to investigate both how model editing affects the general robustness of a model, as well as the robustness of the specific behavior targeted by the edit. We find that edits tend to reduce general robustness, but that the degree of degradation depends on the editing algorithm and layers chosen. Motivated by these observations we introduce a new model editing algorithm, 1-layer interpolation (1-LI), which uses weight-space interpolation to navigate the trade-off between editing task accuracy and general robustness.

Figures (13)

• Figure 1: 1-layer interpolation displays stand-out effective robustness after fine-tuning on an editing task. See Section \ref{['sec:results-1li']}. ImageNet-R vs ImageNet (top) and task accuracy (bottom) along line segments $(1-\alpha) W + \alpha W'$ between original and edited weights $W$ and $W'$ of CLIP models edited on various tasks, where different points correspond to different interpolation parameters $\alpha$.
• Figure 2: OOD performance of edited models. (left) Editing validation accuracies of VGG16 and ViT models on the synthetic concept-style task. (middle) Editing task OOD penalties \ref{['eq:edit-ood-penalty']} of the edited models in (a). (right) original task OOD penalties on ImageNet-C \ref{['eq:orig-ood-penalty']} of VGG16 models edited using the synthetic concept-style task, using the local fine-tuning for ouput collision (\ref{['item:loc-ft-oc']}) and rewriting (\ref{['item:rewrit']}) methods.
• Figure 3: Linear interpolation between original and edited weights $W$ and $W'$ for edited VGG16 models the vehicles-on-snow dataset. Each curve corresponds to a an editing layer, and each point on that curve corresponds to evaluation of a model with weights $(1-\alpha)W + \alpha W'$ for some $t \in [0, 1]$. (left): local fine-tuning for output collision. (right): direct rank one editing.
• Figure 4: Linear interpolation between original and edited weights $W$ and $W'$ of CLIP models edited on various tasks. Each curve corresponds to a an editing layer (or to global fine-tuning), and each point on that curve corresponds to evaluation of a model with weights $(1-\alpha)W + \alpha W'$ for some $\alpha \in [0, 1]$.
• Figure 5: ImageNet-C accuracy along line segments $(1-\alpha) W + \alpha W'$ between original and edited weights $W$ and $W'$ of CLIP models edited on various tasks, plotted with respect to the interpolation parameter $\alpha$.