finding
active
finding:positional-perturbation-raises-extraction-length-for-k-fac-from-6-6-to-13-3-tokens-and-bsn-4-6-to-10-less-than-prior-methods-13-to-16Positional perturbation raises extraction length for K-FAC from 6.6 to 13.3 tokens and BSN 4.6 to 10, less than prior methods (+13 to +16)
Robustness.
Source paper
extracted_from(2025) · Merullo, Jack · Vatsavaya, Srihita · Bushnaq, Lucius · Lewis, Owen
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- K-FAC suppresses the broadest range of unseen memorized text without requiring labeled forget datasupportsPractical advantage over BSN.
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.
- K-FAC edit reduces OLMo-7B Dolma strict accuracy from 99.9% to 3.4% (BSN 6.0%, SVD 3.0%)finding0.741Table 1.
- Better transfer to unseen memorized data.
- K-FAC eigenvector bands show sorted activation strength top10%>10-25%>25-50%>bottom50%, unlike SVDfinding0.737Supports curvature basis interpretability.
- Regularization effect.
- Block and Grassmannian featurizers keep utilization ≈0.6–0.66 at k=16, while Group Lasso falls to ≈0.17.finding0.729Per-feature occupancy.
- Blocking attention to query token in layers 11–16 drops accuracy from 98% to 37%; with next token 8%finding0.726Attention knockout.
- Justifies the choice of response tokens for persona vector extraction in the pipeline
- Contrasts with temporal permutation where Span Representation dominates; suggests spatio permutation reveals different dynamics.