hypothesis
active
hypothesis:the-answer-selection-head-is-monosemanticThe answer selection head is monosemantic
Best guess but not fully convinced; lacks crisp text description to test.
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (2)
finding
- Supports monosemantic reading.
- PCA uninformative here.
Questions (1)
question
- Raised by disjoint feature sets.
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.
- Head selecting relevant content; generalizes beyond multiple choice.
- Will a head perform 'the same algorithm' across prompts, or are individual heads polysemantic?question0.799Head polysemanticity.
- None of the top-activating examples for the answer selection head are multiple choice questionsfinding0.790Shows broader behavior.
- Interpretability property where a latent feature represents a single semantic concept; benchmarked across architectures.
- QK attribution for MC heads.
- A head can implement multiple behaviors via high-rank output without superposition.
- MC interpretation.
- Theoretical point.