Selecting the highest-norm columns of the activation matrix inside each matrix product preserves quality at reduced compute, with attention far more reducible than MLP layers.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Reduced Matrix Multiplication: Input-Adaptive Matrix-Product Reduction for LLM Inference
Selecting the highest-norm columns of the activation matrix inside each matrix product preserves quality at reduced compute, with attention far more reducible than MLP layers.