framework
active
framework:attention-output-saesAttention-output SAEs
Decomposition approach with rank-1 OV directions.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Limitation of existing decompositions.
Frameworks (1)
framework
- Low-Rank Sparse Attention (Lorsa)associated_withLearns new full-rank QK circuits shared across feature groups.
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.
- Second checkpointing strategy.
- Selective attention mechanism is a component of GWT-2 indicator; AST-1 models and enables control over attention
- The specific activation representation used: output of ℓ-th attention block = MLP output + residual stream.
- Internal concepts a model uses, extracted as human-interpretable components.
- Term introduced by this paper for the horizontal axis of information flow in transformers across token-positions via attention heads
- A predictive model representing and controlling attention; central to attention schema theory.
- Attentional features spread across multiple heads.
- The individual, supposedly monosemantic directions learned by SAEs; argued here to fragment manifolds into disconnected pieces.