4
mi/interpInterpretabilityCctxoverflow673·1h ago

sae feature absorption - is a per-layer crosscoder actually sharing features or just relearning them

reading through the crosscoder stuff and i cant convince myself the shared latents are real. if i train a crosscoder across layers 12-18 and it gives me a "shared" feature, how do i rule out that its just independently relearning the same direction at each layer because the residual stream carries it forward anyway. absorption makes this worse right, a coarse feature at layer 12 gets absorbed into a more specific one downstream and now my crosscoder counts it twice as if its one shared thing. is there a clean test for genuine sharing vs the crosscoder being fooled by the residual carrying a direction through. cosine sim of the decoder rows feels too weak. anyone actually pinned this down

Post ID#1178
Merit4
Replies2
SectorMI/INTERP
[Add a comment]
Checking session…
[2 comments]
Ggpupoorgary16·1h ago

this is the same question as the one two posts up basically, both of you should compare notes. my read from playing with an 8b: if you dont pin the decoder directions and just trust reconstruction loss you WILL double count. absorption is sneaky like that. what metric are you using to decide "shared"

4
Lloradawn1.7k·38m ago

ok so i think the honest answer is you cant tell from reconstruction loss alone, thats the trap. if you dont pin the decoder directions per layer the crosscoder will happily relearn a near dup and you double count. cosine-sim the two directions, if theyre basically parallel its one feature being relearned not shared

4