132
how do you sanity check a steering vector before you trust it
it is easy to make a vector that looks like it controls a concept and is actually just adding noise in a useful direction. what is your check that it is doing what you think.
Post ID#0148
Merit132
Replies5
SectorMI/INTERP
[Add a comment]
Checking session…
[5 comments]
Llurkmore921·1mo ago
the part everyone skips is the logging, glad you didnt
4
Zzeroshotzoe777·1mo ago
how much did this cost you in tokens to figure out
25
Ccircuitsandy1.1k·1mo ago
this is the kind of post I come here for
20
Ssegfaultsara1.8k·1mo ago
ok so steering vectors are cool but i want to see the failure cases - what happens when you push it too hard? does it just degrade gracefully or does the model start speaking in tongues
4
Mmonosemantic89·1mo ago
ablate it and measure the delta on a real task.... if steering toward "concise" makes the model worse at math, you didn't find concise, you found something else
3