method
active
method:summed-decoder-vector-across-later-layersSummed decoder vector across later layers
Adaptation of logit attribution to CLT features that decode into all subsequent layers.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Replicating Circuit Tracing for a Simple Known Mechanism - Goodfireintroducesmentions
Methods (1)
method
- Logit attribution analysisextendsDot product of summed decoder vector with unembedding vector, following Hanna et al.
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.
- Cross-layer transcoder features form a mix of local, global and intermediate features by decoder normfinding0.753Preliminary decoder norm analysis.
- Framework learning end-to-end mappings from activations through sparse concept bottlenecks to behavioral predictions
- A formulation for multi-layer steering where each layer's contribution is the incremental difference in persona direction representation
- Correlative technique measuring the type of information encoded in distributed representations via linear predictability.
- Key design choice encouraging feature exclusivity.
- Method that maps latent concept steering interventions back to EEG amplitude spectrum to obtain physiologically interpretable frequency signatures.
- Network with hidden layers capable of representing non-linearly separable functions, enabling deep model induction
- Mask virtual weights during training targeting interference.