Back to feed
arXiv cs.LG·

Geometric Erasure by Contrastive Velocity Matching in Rectified Flows

Signal
72
Hype
28
In three linesGEM is a concept erasure framework for Rectified Flow Transformers. It bridges trajectory-based unlearning (Generative Flow Networks) and teacher-guided erasure, using geometric guidance signals to suppress unwanted concepts while preserving benign generation and preventing harmful content synthesis.
Read source
Your take?
AI safetyAlignmentPapersImage generationVideo generation

Summary generated by Claude — human-verified