finding
active
finding:rare-relations-like-company-ceo-drop-78-under-k-fac-edit-while-frequent-ones-barely-changeRare relations like Company-CEO drop 78% under K-FAC edit while frequent ones barely change
Frequency modulates brittleness.
Source paper
extracted_from(2025) · Merullo, Jack · Vatsavaya, Srihita · Bushnaq, Lucius · Lewis, Owen
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Abstract thesis.
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.
- Better transfer to unseen memorized data.
- K-FAC edit reduces OLMo-7B Dolma strict accuracy from 99.9% to 3.4% (BSN 6.0%, SVD 3.0%)finding0.725Table 1.
- K-FAC edit on ViT-Base drops memorized train accuracy from 81.6% to 3.5%, GT accuracy to 66.5% (SVD 58.9%)finding0.721Table 2.
- Regularization effect.
- MMLU-Pro Math drops from 23.4% to 15.8% under K-FAC while computer science stays 26.1% to 26.3%finding0.715Arithmetic specifically hurt.
- Logic sometimes improves.
- Shows that explicit labels without latent steering fail to generalize to situational cues
- Fig. 3 and Fig. 19.