concept
active
concept:head-polysemanticity-without-superpositionHead polysemanticity without superposition
A head can implement multiple behaviors via high-rank output without superposition.
Neighborhood — ranked by edge-count
Papers (1)
paper
- HeadVis: An Interactive Tool For Investigating Attention Headsintroducesmentions
Concepts (2)
concept
- Attention biologyassociated_withStudy of what attentional features and head behaviors exist.
- MLP neuron superpositionanalogous_toMLP neuron polysemanticity requires superposition, unlike heads.
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.
- Neurons that respond to multiple unrelated concepts, limiting interpretability.
- Most pragmatically important question.
- Will a head perform 'the same algorithm' across prompts, or are individual heads polysemantic?question0.802Head polysemanticity.
- Interpretability property where a latent feature represents a single semantic concept; benchmarked across architectures.
- Major prediction checked against InceptionV1.
- Best guess but not fully convinced; lacks crisp text description to test.
- Theoretical point.
- Author is fairly confident.