claim
active
claim:a-head-s-behavior-on-the-full-distribution-rarely-matches-what-a-narrow-task-suggestsA head's behavior on the full distribution rarely matches what a narrow task suggests
Theme across prior and own work.
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (2)
finding
- Identified from attention patterns in Haiku 3.5.
- None of the top-activating examples for the answer selection head are multiple choice questionssupportsShows broader behavior.
Quotes (1)
quote
- Central theme.
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.
- Prediction for future decomposition.
- Does the answer selection head really implement one behavior, or many behaviors that share a common template?question0.792Raised by disjoint feature sets.
- Head-level causal evidence.
- Will a head perform 'the same algorithm' across prompts, or are individual heads polysemantic?question0.787Head polysemanticity.
- MC interpretation.
- Claim supported by VPD's recovery of cross-head attention subcomponents, noted in footnote.
- Data-induced competition over resources.
- Empirical observation from examining expanded OV/QK matrices; approximately 10 out of 12 heads show significant copying