Merging subtoken hidden states of the same word into one vector reduces FLOPs by up to 19% across six code models while keeping downstream F1 and CodeBLEU roughly unchanged.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.SE 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
On the Effect of Token Merging on Pre-trained Models for Code
Merging subtoken hidden states of the same word into one vector reduces FLOPs by up to 19% across six code models while keeping downstream F1 and CodeBLEU roughly unchanged.