SAGE: Training-Free Semantic Evidence Composition for Edge-Cloud Inference under Hard Uplink Budgets

arXiv cs.LG / 4/22/2026

📰 NewsDeveloper Stack & InfrastructureModels & Research

共有:

Key Points

The paper argues that collaborative edge-cloud inference under strict per-request uplink bit budgets cannot rely on attention-based importance selection alone, since that approach has inherent limitations.
It shows that substituting high-importance segments with lower-importance but complementary ones improves server accuracy, indicating that coverage diversity—not individual importance—drives performance.
It further finds that spatially uniform evidence selection without content information can reach competitive accuracy at moderate budgets, highlighting the independent value of spatial coverage.
Based on these insights, the authors propose SAGE (Semantic Attention-Guided Evidence), a training-free method that blends importance filtering with embedding-diversity sampling, achieving about 93% of server-side ceiling accuracy while sending less than half the evidence units on ImageNet-1K.
The results demonstrate a substantial improvement over importance-only composition in offloaded accuracy versus transmitted data volume trade-offs.

Abstract

Edge-cloud hybrid inference offloads difficult inputs to a powerful remote model, but the uplink channel imposes hard per-request constraints on the number of bits that can be transmitted. We show that selecting transmitted content based solely on attention-based importance, the standard approach in collaborative inference, is inherently limited under hard budgets. Two findings support this claim. First, replacing high-importance units with low-importance but complementary ones improves server accuracy. This shows that what matters is not individual importance but how well the transmitted set covers diverse aspects of the input. Second, spatially uniform selection without any content information achieves competitive accuracy at moderate budgets. This confirms that spatial coverage alone carries independent value. Based on this analysis, we propose SAGE (Semantic Attention-Guided Evidence), a principled, training-free method that combines importance filtering with embedding-diversity sampling. SAGE achieves 93% of the server ceiling in offloaded accuracy while transmitting fewer than half of the available evidence units on ImageNet-1K, substantially outperforming importance-only composition.