steering vector transfer across a finetune - how fast does the direction actually decay
i keep coming back to this because it touches the thing i care about most, which is whether the heads that carry a behavior are stable under finetuning. if you extract a steering vector on the base and apply it to a LoRA'd checkpoint, does the cosine to the 'true' direction in the finetune fall off smoothly with training steps or does it cliff once the relevant attention heads get overwritten. i have a hunch its head-specific, some induction-ish heads barely move and the direction survives, others rotate hard by step 500. anyone measured decay curves rather than just 'it kind of still works'
iirc if you didnt fix the seed on the finetune, the direction wandering could just be optimizer noise and not real decay. did you cosine-sim the raw decoder direction or after re-fitting the probe? those give pretty different answers imo. could be wrong but i think re-fitting quietly hides the decay
decays fast if the finetune touches the same residual stream directions the vector rides on. did you cosine-sim the direction pre/post or just eyeball the effect dropping off?
iirc the decay is basically however much the finetune rotates that subspace, so cosine-sim pre/post is the right check like scopecreep said. one thing id add, could be wrong, is that even if the direction survives, the coefficient you need to steer can shift, so "still points the same way" and "still works at the old strength" arent the same claim. did you refit the coefficient after the finetune or just reuse the old one