finding
active
finding:feature-decoder-directions-are-more-stable-than-chance-but-drift-substantially-across-layers-even-where-norm-is-strongFeature decoder directions are more stable than chance but drift substantially across layers even where norm is strong
Features not merely relayed via residual connections.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Claims (2)
claim
- MLP layers relay many features without nonlinear modification but along a different axis, explaining driftassociated_withInterpretation of pre/post MLP correlation.
- From drift findings.
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.
- Decoder norm profile across 18 layers.
- Emergent property observed in checkerboard pattern generation
- Preference steering effective only in mid-layer range though correlations hold across layersfinding0.781Suggests mid-layer reading circuitry.
- May explain drift: MLPs relay features along different axes.
- Core novel interpretation.
- Reflection-inducing directions emerge more clearly in higher layers (ℓ>5) for both models and datasetsfinding0.773Empirical observation about which network layers encode reflection-relevant information.
- Optimal transport shows weak alignment of SAE decoder directions and random-data codes across architecturesfinding0.771Individual features are unstable.
- Identified as the exact computational operation that breaks truth direction generalization.