claim
active
claim:there-is-significant-redundant-linearly-correlated-structure-across-layers-interpreted-as-cross-layer-featuresThere is significant redundant, linearly correlated structure across layers interpreted as cross-layer features
Interpretation of eval loss result.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Findings (1)
finding
- Indicates redundant structure across layers.
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.
- Consistent but alternative explanations like amplification exist.
- From drift findings.
- Explicitly identified research gap: anecdotal evidence exists but rigorous characterization is absent
- Representation of features spread across multiple layers, complicating dictionary learning.
- Cross-layer transcoder features form a mix of local, global and intermediate features by decoder normfinding0.780Preliminary decoder norm analysis.
- Decoder norm profile across 18 layers.
- Motivating example with 13 vs 3 features.
- A random head in the same layer shows almost no structure between count and width probesfinding0.769Control for twisting.