Multi-head softmax attention can be provably recovered from black-box scalar value queries in O(H d^2) queries without orthogonal-subspace assumptions, using rational interpolation followed by head matching.
International Conference on Learning Representations , year =
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
ACCEPT 1representative citing papers
citing papers explorer
-
Provably Learning Multi-Head Attention with Queries
Multi-head softmax attention can be provably recovered from black-box scalar value queries in O(H d^2) queries without orthogonal-subspace assumptions, using rational interpolation followed by head matching.