finding
active
finding:inhibiting-query-side-opposite-features-reduces-grand-prediction-and-induces-synonyms-like-peu-and-faible-inhibiting-key-side-adjective-features-has-lesser-effectInhibiting query-side opposite features reduces 'grand' prediction and induces synonyms like peu and faible; inhibiting key-side adjective features has lesser effect
Steering validation for antonym mechanism.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Antonym interpretation.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Perturbation validation.
- QK attribution of antonym heads.
- Explains why cities+neg_cities and larger_than+smaller_than training sets yield better OOD accuracy
- QK attribution for MC heads.
- English is mechanistically the default output.
- Limitation of single-prompt selection.
- Implicit modulation.
- Inhibiting RM bias features reduces 9-1-1 mentions to 6% and voting to 0%; baseline 42% 9-1-1, 56% voting.finding0.747Law-schools example.