A masked, single-weight-matrix gating scheme, MGLU, matches Gated Linear Unit accuracy on tested language tasks while reducing per-token memory reads by up to 47 percent.
Table 16: MGLU latency and speed-ups
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Masked Gated Linear Unit
A masked, single-weight-matrix gating scheme, MGLU, matches Gated Linear Unit accuracy on tested language tasks while reducing per-token memory reads by up to 47 percent.