EraseFlow: Learning Concept Erasure Policies via GFlowNet-Driven Alignment

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Kusumba, Abhiram, Patel, Maitreya, Min, Kyle, Kim, Changhoon, Baral, Chitta, Yang, Yezhou
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866917068977209344
author Kusumba, Abhiram
Patel, Maitreya
Min, Kyle
Kim, Changhoon
Baral, Chitta
Yang, Yezhou
author_facet Kusumba, Abhiram
Patel, Maitreya
Min, Kyle
Kim, Changhoon
Baral, Chitta
Yang, Yezhou
contents Erasing harmful or proprietary concepts from powerful text to image generators is an emerging safety requirement, yet current "concept erasure" techniques either collapse image quality, rely on brittle adversarial losses, or demand prohibitive retraining cycles. We trace these limitations to a myopic view of the denoising trajectories that govern diffusion based generation. We introduce EraseFlow, the first framework that casts concept unlearning as exploration in the space of denoising paths and optimizes it with GFlowNets equipped with the trajectory balance objective. By sampling entire trajectories rather than single end states, EraseFlow learns a stochastic policy that steers generation away from target concepts while preserving the model's prior. EraseFlow eliminates the need for carefully crafted reward models and by doing this, it generalizes effectively to unseen concepts and avoids hackable rewards while improving the performance. Extensive empirical results demonstrate that EraseFlow outperforms existing baselines and achieves an optimal trade off between performance and prior preservation.
format Preprint
id arxiv_https___arxiv_org_abs_2511_00804
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle EraseFlow: Learning Concept Erasure Policies via GFlowNet-Driven Alignment
Kusumba, Abhiram
Patel, Maitreya
Min, Kyle
Kim, Changhoon
Baral, Chitta
Yang, Yezhou
Machine Learning
Computer Vision and Pattern Recognition
Erasing harmful or proprietary concepts from powerful text to image generators is an emerging safety requirement, yet current "concept erasure" techniques either collapse image quality, rely on brittle adversarial losses, or demand prohibitive retraining cycles. We trace these limitations to a myopic view of the denoising trajectories that govern diffusion based generation. We introduce EraseFlow, the first framework that casts concept unlearning as exploration in the space of denoising paths and optimizes it with GFlowNets equipped with the trajectory balance objective. By sampling entire trajectories rather than single end states, EraseFlow learns a stochastic policy that steers generation away from target concepts while preserving the model's prior. EraseFlow eliminates the need for carefully crafted reward models and by doing this, it generalizes effectively to unseen concepts and avoids hackable rewards while improving the performance. Extensive empirical results demonstrate that EraseFlow outperforms existing baselines and achieves an optimal trade off between performance and prior preservation.
title EraseFlow: Learning Concept Erasure Policies via GFlowNet-Driven Alignment
topic Machine Learning
Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2511.00804