finding
active
finding:preliminary-analyses-suggest-qk-attribution-matrices-per-head-and-position-pair-are-approximately-low-rankPreliminary analyses suggest QK attribution matrices per head and position pair are approximately low-rank
Motivates distillation.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Papers (1)
paper
Questions (1)
question
- Can QK attribution information be distilled or simplified?answered_bysupportsFuture work.
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.
- In the Aunt Mary Sue prompt, QK attributions matrix has rank at least 2 with no strong cross termsfinding0.813Parallel independent QK interactions.
- Central claim.
- Quadratic scaling.
- Core method decomposing attention scores as bilinear sums of query-key feature pairs.
- Reframing observation: the canonical K/Q/V decomposition is computationally convenient but not the most interpretable representation
- Quantifies method limitations.
- Writes attention scores/head output as sums of feature interactions; core HeadVis capability.
- Feature attribution (gradient-based) correlates 0.8 with ablation effects on the 'John' and 'Kobe' examples.finding0.758Validation of attribution as a fast proxy for causal importance.