Rescale outside the span, renormalize
One attention row in one steered head. Click tokens to define the highlighted span. Illustrative weights The operation itself follows the paper.
outside span: score × α → renormalize to sum 1 ⇒ span relatively ×(1/α). Paper fixes α = 0.01.
Why multiply, not add?
Shares inside the span only. Illustrative Adding a large constant flattens differences; scaling the outside keeps them.
Four ways to control a model
Click a box. PASTA is none of the other three.
Where a head gets edited
ITI is the closest relative: it also picks a sparse set of heads.
Profile once per model
Steering all heads performs worse than zero-shot; single heads or layers swing wildly.
LLaMA-7B: 32 layers × 32 heads = 1024
Per-head scores are a synthetic pattern Illustrative — not the paper's measurements. Slide k to see a top-k cut.
Hover a cell
Reported: head-set sizes and selection strategy
Task sets end up at 53, 86, 86, 86 heads (task order not given in episode). Strategy scores are averages.
Reported results (LLaMA-7B unless noted)
All values are percentages, higher is better. "~" marks approximate figures from the episode.
Alpha: forgiving, except zero
Stable from roughly 0.05 down to 1e-4.
Fluency guard (bigram + trigram entropy)
Outputs below 3.0 are discarded. Appendix D failure: repeated "They are a photographer."
How held-out is "task-agnostic"?
Tested vs not shown
Per the hosts' discussion of the paper.