Adding learned per-head routing over previous layers' key-value buffers to a Transformer reduces representation collapse, lowers language modeling loss, and improves synthetic arithmetic and planning accuracy.
Cross-layer retrospective retrieving via layer attention
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
-
You Do Not Fully Utilize Transformer's Representation Capacity
Adding learned per-head routing over previous layers' key-value buffers to a Transformer reduces representation collapse, lowers language modeling loss, and improves synthetic arithmetic and planning accuracy.