HSMLA extends EfficientViT-style multi-scale linear attention with a learned gate that sends a fraction of blocks to local softmax refinement, reporting up to 4.2x speedups with matched or better accuracy.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
HSMLA: Hierarchical Softmax Multi-scale Linear Attention for Efficient Vision Transformers
HSMLA extends EfficientViT-style multi-scale linear attention with a learned gate that sends a fraction of blocks to local softmax refinement, reporting up to 4.2x speedups with matched or better accuracy.