Skip to content
arXiv· arXiv:2608.13478· Ashok Urlana, L. D. M. S. Sai Teja, Vivek Hruday Kavuri, Ponnurangam Kumaraguru·Indexed by this site · Originally published

MapRoute++: Agent-Guided Semantic Routing for Visual Concept Unlearning

MapRoute++: Surrogate-Guided Semantic Routing for Visual Concept Unlearning

AI summary

The authors propose MapRoute++, which on the official benchmark of the Genμ 2.0 Challenge Task 3 (visual concept forgetting), evaluated with the ERR metric on Stable Diffusion v1.4, surpasses the SOTA baseline by an average of 12.1% across five concept categories. Building on MapRoute, the method introduces task-specific training objectives and richer concept representations, and uses semantic routing to select the corresponding mapper for different concepts, more robustly removing target concepts while preserving irrelevant and semantically adjacent concepts.

Read originalarxiv.org