aiqbench analysis
The paper studies how the choice of "anchor" concept affects erasure in text-to-image diffusion models. It finds that concepts with more dispersed causal effects are harder to erase robustly, and it proposes SELECT, which picks counterfactual redirection targets from semantically close but mutually exclusive concepts. The reported gains are in suppression and preservation trade-offs on object, celebrity, artist-style, and NSFW benchmarks. Readers building or auditing generative image systems should note that the evidence is from academic benchmarks, so robustness against adversarial re-elicitation still needs independent testing.
SELECT: Semantic Erasure Boundary Construction with Sibling Exclusive Counterfactuals
arXiv:2510.16342v2 Announce Type: replace Abstract: Concept erasure aims to suppress undesirable concepts in text-to-image diffusion models while preserving their generative capabilities. Existing methods often redirect target concepts toward predefined semantic anchors, but the influence of anchor selection on the scope of concept erasure remains insufficiently understood. Through causal tracing, we reveal that concepts with more dispersed causal effects tend to exhibit higher anchor sensitivity and lower erasure robustness. Controlled anchor interventions demonstrate that anchor selection affects target suppression and collateral damage to surrounding semantic concepts. These findings motivate us to formulate anchor selection as semantic erasure boundary construction. We propose SELECT (Selective Erasure via Local Exclusive Counterfactual Targets), a plug-and-play framework based on Sibling Exclusive Concepts (SECs), which are semantically close to the target but mutually exclusive in its defining attributes. SELECT jointly constructs a counterfactual redirection endpoint and a local retention set to balance target suppression and semantic preservation. SELECT filters candidates using contextual activation, and uses semantic coherence to identify context compatible redirection endpoints, while the remaining SECs serve as local preservation constraints. Experiments on object, celebrity, artist-style, and NSFW erasure show that SELECT strengthens target suppression over fixed general anchors and achieves suppression-preservation trade-offs, while reducing concept re-emergence and erosion.
Abstract:Concept erasure aims to suppress undesirable concepts in text-to-image diffusion models while preserving their generative capabilities. Existing methods often redirect target concepts toward predefined semantic anchors, but the influence of anchor selection on the scope of concept erasure remains insufficiently understood. Through causal tracing, we reveal that concepts with more dispersed causal effects tend to exhibit higher anchor sensitivity and lower erasure robustness. Controlled anchor interventions demonstrate that anchor selection affects target suppression and collateral damage to surrounding semantic concepts. These findings motivate us to formulate anchor selection as semantic erasure boundary construction. We propose SELECT (Selective Erasure via Local Exclusive Counterfactual Targets), a plug-and-play framework based on Sibling Exclusive Concepts (SECs), which are semantically close to the target but mutually exclusive in its defining attributes. SELECT jointly constructs a counterfactual redirection endpoint and a local retention set to balance target suppression and semantic preservation. SELECT filters candidates using contextual activation, and uses semantic coherence to identify context compatible redirection endpoints, while the remaining SECs serve as local preservation constraints. Experiments on object, celebrity, artist-style, and NSFW erasure show that SELECT strengthens target suppression over fixed general anchors and achieves suppression-preservation trade-offs, while reducing concept re-emergence and erosion.
Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV)
arXiv:2510.16342 [cs.AI]
(or arXiv:2510.16342v2 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2510.16342 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 296 words
Related coverage
G2VD: Generalizable AI-Generated Video Detection via Counterfactual Intervention and Causal Disentanglement
arXiv · Jul 31, 2026
Backtracking Direct, Indirect and Total Effects
arXiv · Oct 22, 2026
StateGuard: Analytical-State Management with Validity-Aware Intervention for Long-Horizon Data Agents
arXiv · Sep 29, 2026
Mitigating Bias in Large Vision-Language Models via Counterfactual Ensemble Decoding
arXiv · Aug 25, 2026