Tree-structured feed-forward layers with hard hierarchical routing deliver dynamic sparsity in large transformers, matching dense baselines at under 5% activation for autoregressive language modeling and downstream question answering.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Dynamic sparsity in tree-structured feed-forward layers at scale
Tree-structured feed-forward layers with hard hierarchical routing deliver dynamic sparsity in large transformers, matching dense baselines at under 5% activation for autoregressive language modeling and downstream question answering.