claim
active
claim:an-attentional-feature-would-look-roughly-like-an-attention-headAn attentional feature would look roughly like an attention head
Expectation for decomposition.
Neighborhood — ranked by edge-count
Papers (1)
paper
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.
- Proposed attention analog of sparse monosemantic features for decomposing attention layers.
- Suspected but no confirmed example in real LLM.
- Features whose QK/OV are high rank, hard to interpret.
- Transformer attention heads that could be recruited to extract different kinds of information (text vs. thoughts).
- Higher-order generative model variables modulating perceptual precision; subject to policy selection (deliberate shifts of attention).
- Prediction for future decomposition.
- Mathematical equivalence enabling independent analysis of each attention head
- Variable aperture of attention, ranging from collapsed to expanded; human and model analogs.