MUtE: A Dual Framework for Concept Erasure and Counterfactual Interventions

Excerpt

arXiv:2609.11253v1 Announce Type: cross Abstract: Erasing concept-specific information from representations has been proven useful for mitigating bias or interpreting model decisions. The joint objective is to transform the original representations such that the target concept becomes unpredictable, while maximally preserving concept-unrelated information. In this work, we revisit the optimal bounds of concept erasure to derive a novel class of erasure functions that naturally induce a determini