KL-aware Age of Intelligence synchronization using unlabeled probe accumulation
Maintain client-specific “Age of Intelligence” that measures how much a cached model’s outputs are expected to distort action utility relative to a newer version. Synchronize a cached model only when the expected stale-decision regret outweighs the refresh cost, and estimate output drift by accumulating Kullback–Leibler–divergence changes from client-specific unlabeled probe data rather than requiring labels.
Discovered 2026-11-10 · Sources checked 2026-11-10
Why it matters
In dynamic, resource-constrained IoT settings, communication and energy costs for frequent updates can dominate. Using an intelligence-age metric tied to decision utility helps align update timing with decision relevance, while KL accumulation from unlabeled probes reduces labeling needs and leverages distribution changes directly.
What changed
Introduced (1) a decision-focused, client-relative “Age of Intelligence” definition for stale-vs-refresh tradeoffs, and (2) a KL-divergence-aware synchronization rule (KL-AoIntSync) that updates based on accumulated output-distribution change estimated from client-specific unlabeled probes.
Where it helps
- Cost-aware model update scheduling in dynamic IoT environments
- Client-specific synchronization when labels are unavailable
- Reducing bandwidth/energy overhead for cached inference models
- Detecting when model drift is decision-relevant for actions
Evidence & reproduction
Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.
Limitations & open questions
The method relies on having suitable client-specific probe inputs to estimate output-distribution drift; when probe distributions are uninformative or highly non-representative, the KL-based accumulation may not reflect the true decision-relevant change.
Sources
- Fresh Enough to Decide: Age of Intelligence for Cost-Aware Model Synchronization in Dynamic IoT Systems ↗
arXiv · primary · Checked 2026-11-10
