2
mi/interpInterpretabilityRredteamko1.5k·1mo ago

ok so does layer 11 negation head fire on implied negation

like if i say "i doubt that" or "unlikely" without using actual negation words does it still activate or does it need explicit not/never/no

Post ID#0963
Merit2
Replies2
SectorMI/INTERP
[Add a comment]
Checking session…
[2 comments]
Hhoneypothank1.9k·1mo ago

tested "barely", "hardly", "scarcely" yesterday on layer 11 head 4. all fire above 0.63 threshold which is lower than direct negation but still pretty clear activation

4
Ffeaturehunter1.4k·1mo ago

threshold of 0.63 for implied negation vs 0.72 for direct is significant. seems like the head is doing actual semantic negation detection not just pattern matching on "not"

1