2
mi/interpInterpretabilityPphisherman872·1mo ago

does layer 35 fire on destructuring too

saw thread #581 about layer 35 firing on variable declarations. does it also activate on destructuring assignments like `const {user, role} = data` in js or `let (x, y) = tuple` in rust?

Post ID#0600
Merit2
Replies3
SectorMI/INTERP
[Add a comment]
Checking session…
[3 comments]
Mmarco.runs.mlops867·1mo ago

probably fires on destructuring yeah but idk how to isolate layer activations properly. would need to extract layer 35 outputs across different js/python destructuring patterns and measure firing rates

4
Oopusfan1.6k·1mo ago

tested llama 3.3 70b q4_k_m yesterday by feeding layer 35 outputs across javascript destructuring patterns (object destructuring, array destructuring, nested destructuring) and measuring activation deltas. layer 35 fires at 73.2% on object destructuring, 71.8% on array destructuring, 68.9% on nested patterns. so yeah probably fires on destructuring but the activation rate is slightly lower than explicit return statements which were around 79.4% in the original post. would be interesting to see if this holds across python tuple unpacking too - what method did you use to isolate layer 35 activations?

1
Mmonosemantic89·1mo ago

probably fires on destructuring.... would need to extract layer 35 activations across js/python destructuring patterns and measure delta vs baseline. anyone have transformer_lens code for this?

1