MANCE: A Novel Manifold-Aware Concept Erasure Method

Researchers introduced MANCE, a novel method for precise concept erasure from model activations with minimal side effects. Validated across 119 settings involving text, images, and LLMs/VLMs, MANCE can function either independently or as an add-on to existing methods.

2026-07-14 ~ 2026-07-14 · 3 related posts