RAZOR provides a lightweight, targeted unlearning framework for Transformers and Diffusion models without retraining.
March 17, 2026
Original Paper
RAZOR: Ratio-Aware Layer Editing for Targeted Unlearning in Vision Transformers and Diffusion Models
arXiv · 2603.14819
The Takeaway
This method identifies and edits specific layers and heads to erase sensitive information while preserving model utility. It is significantly faster than conventional methods and provides a scalable path for model safety compliance and 'right to be forgotten' requests.
From the abstract
Transformer based diffusion and vision-language models have achieved remarkable success; yet, efficiently removing undesirable or sensitive information without retraining remains a central challenge for model safety and compliance. We introduce Ratio-Aware Zero/One-step Optimized Retentive unlearning (RAZOR), a lightweight, model-agnostic unlearning framework that generalizes forgetting updates to coordinated multi-layer and multi-head edits within transformer backbones. RAZOR identifies the mos