Paper proposes I-CARE framework to formalize and study interference in generative machine unlearning

A new paper introduces I-CARE, a methodology designed to formalize interference as a central focus in the study of generative machine unlearning. Unlike prior work that proposes new benchmarks or algorithms, I-CARE provides formal definitions for tasks, metrics, and reporting templates to enable systematic and reproducible analysis of interference across unlearning settings.

The authors argue that interference—defined here as the unintended degradation of semantically related concepts that should have been retained—remains poorly characterized and inconsistently evaluated in current generative unlearning research. By elevating interference to a first-class object of study, the framework seeks to standardize how such effects are measured and reported.

To demonstrate feasibility, the paper presents results using state-of-the-art unlearning algorithms and commonly used datasets. The authors report that I-CARE enables meaningful analysis of interference patterns across multiple unlearning settings, establishing the practical applicability of the framework.

The methodology is explicitly designed to remain valid as models and unlearning algorithms evolve, decoupling long-term scientific insights from transient empirical outcomes. The authors release an open-source implementation of the framework, along with a web-based graphical interface to facilitate exploration of study outcomes without requiring direct interaction with code or specialized data analysis tools.