hypothesis
active
hypothesis:gradual-feature-formation-across-layers-is-evidence-of-cross-layer-superpositionGradual feature formation across layers is evidence of cross-layer superposition
Consistent but alternative explanations like amplification exist.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Findings (1)
finding
- Decoder norm profile across 18 layers.
Concepts (1)
concept
- Representation of features spread across multiple layers, complicating dictionary learning.
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.
- Interpretation of eval loss result.
- Features smeared across layers cannot be fully disentangled by SAE on a single residual stream.
- Adversarial example transfer may be predicted by correlated arrangement of features in superpositionhypothesis0.795Speculative.
- superposition is deeply linked to the ability to identify and enumerate over all features in a modelquote0.791Safety rationale.
- From drift findings.
- Analogy to many ice phases.
- Methodological critique of prior work that fixed a single layer for truth probing.
- Cross-layer transcoder features form a mix of local, global and intermediate features by decoder normfinding0.782Preliminary decoder norm analysis.