claim
active
claim:same-set-suppression-is-a-real-mechanism-of-the-headSame-set suppression is a real mechanism of the head
Supported by geometry and 98% result but not a full account.
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (7)
finding
- QK geometry anti-alignment.
- Systematic test of same-set suppression.
- Custom sentences with odd one out.
- In the residual stream a country and its own cities have positive cosine similarity (mean +0.17)supportsHead flips sign via W_Q and W_K.
- Replacing Ivory Coast with Zambia drops attention to 'a' from 0.87 to 0.08; Lusaka gives 0.10supportsSame-set suppression on real example.
- Hint at suppression.
- Only Zambia is suppressed.
Claims (1)
claim
- Author's caveat.
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 suppresses attention between related items, e.g., country and its own cities.
- Mechanism by which activation of an emotion feature sometimes leads to later suppression of that same featurequestion0.774Identified research gap: the paper observes anti-persistence but has no explanation for it
- Central theme.
- Will a head perform 'the same algorithm' across prompts, or are individual heads polysemantic?question0.738Head polysemanticity.
- Small-to-large transformation is mediated by opposite heads that invert semantic meaninghypothesis0.737Prior hypothesis not confirmed.
- Motivates including Head Cor+Anti condition in steering position comparison
- Theme across prior and own work.
- Mechanism agreement.