question
active
question:can-qk-attribution-information-be-distilled-or-simplifiedCan QK attribution information be distilled or simplified?
Future work.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Preliminary analyses suggest QK attribution matrices per head and position pair are approximately low-rankanswered_bysupportsMotivates distillation.
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.
- Core method decomposing attention scores as bilinear sums of query-key feature pairs.
- Central claim.
- Writes attention scores/head output as sums of feature interactions; core HeadVis capability.
- Quadratic scaling.
- Correlating attribution vectors (feature activation × logit weight of next token) across model pairs to measure functional universality
- QK attribution from Ivory to 'a' has top pair Ivory Coast query feature and Zambian key featurefinding0.759Explains attention edge.
- Quantifies method limitations.
- In the Aunt Mary Sue prompt, QK attributions matrix has rank at least 2 with no strong cross termsfinding0.745Parallel independent QK interactions.