Shuhui Wang

NMKE is a fine-grained framework that uses neuron-level attribution and sparse masking for precise, efficient lifelong knowledge editing.

We propose a sparse autoencoder that maps the semantics of vision and language representations into a unified concept set.