finding
active
finding:stable-features-show-positive-but-high-variance-low-in-absolute-terms-correlation-between-pre-and-post-mlp-directionsStable features show positive but high-variance, low-in-absolute-terms correlation between pre- and post-MLP directions
May explain drift: MLPs relay features along different axes.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Claims (1)
claim
- Interpretation of pre/post MLP correlation.
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.
- Features not merely relayed via residual connections.
- Future-work perspective.
- SAE features are not simply mirroring individual neurons.
- A post-MLP feature for words indicating uniqueness has pre-MLP inputs firing on specific such wordsfinding0.749Abstract features built from specific instances.
- Hypothesis raised in distributive law task analysis
- Demonstrates alignment with Linear Representation Hypothesis: target trait steers approximately linearly with alpha
- Comparative finding ranking intervention sites by coherency preservation
- How can the stability and variance of feature discovery across training runs be improved?question0.743Future work.