finding
active
finding:at-olmo-7b-layer-22-mlp-input-bottom-50-k-fac-eigenvectors-activate-23-1-more-on-memorized-data-top-10-activate-26-more-on-cleanAt OLMo-7B layer 22 MLP input, bottom 50% K-FAC eigenvectors activate 23.1% more on memorized data; top 10% activate 26% more on clean
Disentanglement evidence.
Source paper
extracted_from(2025) · Merullo, Jack · Vatsavaya, Srihita · Bushnaq, Lucius · Lewis, Owen
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- If curvature measures memorization, eigenvector bands will activate differently on memorized vs clean dataassociated_withsupportsSection 4 hypothesis.
Questions (1)
question
- Aggregate analysis motivation.
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.
- Vision disentanglement.
- K-FAC edit reduces OLMo-7B Dolma strict accuracy from 99.9% to 3.4% (BSN 6.0%, SVD 3.0%)finding0.812Table 1.
- Better transfer to unseen memorized data.
- K-FAC eigenvector bands show sorted activation strength top10%>10-25%>25-50%>bottom50%, unlike SVDfinding0.792Supports curvature basis interpretability.
- Coherence comparison.
- Fig. 5.
- Replication at smaller scale.
- Claim about the sparsity and sufficiency of the identified neuron set