Do SAEs only work for morphological traits, or do they discover functional traits?

Like the title says. I expect SAEs to be very good at morphological traits because (as I understand it) they're very visual traits. But functional traits require a broader context. Does a vision model like DINOv2 understand functional traits? Can SAEs identify such traits? What about more biologically informed models like BioCLIP?

I think FishVista has functional trait annotations, so we could use that a test bed.

Start by reviewing the FishVista functional trait annotations and the proposed DINOv2, BioCLIP, and SAE approaches. Design an evaluation that tests whether discovered features correspond to functional traits rather than only morphological ones. Done means reporting comparative results and documenting the methodology and conclusions.

Written by the indexing model from the issue text.