Transformer layer diversity, measured via bias-diversity and information-theoretic decompositions, explains why adding layers improves performance only when layers are diverse, and predicts diminishing returns consistent with scaling laws.
When parts are greater than sums: Individual LLM components can outperform full models
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
other 1
citation-polarity summary
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1roles
other 1polarities
unclear 1representative citing papers
citing papers explorer
-
Diversity of Transformer Layers: One Aspect of Parameter Scaling Laws
Transformer layer diversity, measured via bias-diversity and information-theoretic decompositions, explains why adding layers improves performance only when layers are diverse, and predicts diminishing returns consistent with scaling laws.