framework
active
framework:attentional-featuresAttentional features
Proposed attention analog of sparse monosemantic features for decomposing attention layers.
Neighborhood — ranked by edge-count
Papers (1)
paper
- HeadVis: An Interactive Tool For Investigating Attention Headsintroducesmentions
Methods (1)
method
- Transcodersanalogous_toDecomposition method for activations; VPD is compared against transcoders in sparsity-reconstruction tradeoff.
Hypotheses (1)
hypothesis
- Prediction for future decomposition.
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.
- Expectation for decomposition.
- Features whose QK/OV are high rank, hard to interpret.
- Variable aperture of attention, ranging from collapsed to expanded; human and model analogs.
- Higher-order generative model variables modulating perceptual precision; subject to policy selection (deliberate shifts of attention).
- Selective attention mechanism is a component of GWT-2 indicator; AST-1 models and enables control over attention
- Suspected but no confirmed example in real LLM.
- Interpretable concept representations hypothesized as basic units of computation.
- Process using Q, K, V to compute a heat map over K and weighted sum of V.