Long-sequence transformers blow up because softmax attention cannot fit dense local dependency patterns with its low-rank parameterization, and separating local and global attention heads prevents the resulting logit explosion.
Title resolution pending
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
-
Dense Local Dependencies Induce Attention-Logit Explosion and Training Instability During Long-Sequence Transformer Training
Long-sequence transformers blow up because softmax attention cannot fit dense local dependency patterns with its low-rank parameterization, and separating local and global attention heads prevents the resulting logit explosion.