OV 'Concept Lenses' are said to give much better readouts in a vision-language model than a logit lens

A user says Sheridan Feucht found OV 'Concept Lenses' while studying how language models copy text verbatim versus approximately. The user says the technique produced much better readouts than a logit lens in a vision-language model, and suggests Feucht's earlier findings point to a layered semantic structure in transformer language models.