One-layer self-attention with rational or ReLU post-processing cannot sign-represent parity unless heads times degree scales linearly with input length.
Spectral properties of threshold functions
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Lower bounds for one-layer transformers that compute parity
One-layer self-attention with rational or ReLU post-processing cannot sign-represent parity unless heads times degree scales linearly with input length.