SAGE mixes attention importance with embedding diversity sampling to reach 93% of full server accuracy on ImageNet-1K while sending under half the evidence units, beating pure importance selection.
Not all patches are what you need: Expediting vision transformers via token reorganizations,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
SAGE: Training-Free Semantic Evidence Composition for Edge-Cloud Inference under Hard Uplink Budgets
SAGE mixes attention importance with embedding diversity sampling to reach 93% of full server accuracy on ImageNet-1K while sending under half the evidence units, beating pure importance selection.