ConceptPatchscopes
Patchscopes
An activation-reading method that injects an activation patch from one layer into another prompt, using the model's generation capability to translate the activation into text.
Timeline (last 90 days)
Oct 12
The activation-reading tools evaluated are divided into single-token readers and multi-token readers: the single-token side includes Logit lens, J-lens, R-lens, and Tuned lens; the multi-token side includes Natural Language Autoencoders, Sparse Autoencoders, Patchscopes, Template lens, and Oracle Lens.
Unverified50%