method
active
method:saes-on-attention-layer-outputsSAEs on attention layer outputs
Second checkpointing strategy.
Neighborhood — ranked by edge-count
Papers (1)
paper
Methods (1)
method
- Strategy forcing edges to be mediated only by head paths of length 1 to avoid exponential path growth.
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.
- Decomposition approach with rank-1 OV directions.
- Out-of-distribution generalization of SAE features.
- Identification of algorithms implemented in attention layers, distributed across attention headsfinding0.742VPD successfully recovered interpretable attention algorithms (previous-token behavior, syntax-boundary routing) in weight space without requiring manual decomposition across heads.
- Extension of mechanistic interpretability findings to the metacognitive domain
- Term introduced by this paper for the horizontal axis of information flow in transformers across token-positions via attention heads
- The specific activation representation used: output of ℓ-th attention block = MLP output + residual stream.
- Interpretability tool underlying top-feature, details, steering tools.
- The first toy model analyzed; shown to implement an ensemble of bigram and skip-trigram models readable directly from weights