RAG Framework Boosts LEO Satellite Networks by 30% in Efficiency

A team of researchers from multiple institutions has introduced a novel framework to address resource fragmentation in Low Earth Orbit (LEO) satellite networks. The paper, submitted to arXiv on June 13, 2026, presents a RAG-enhanced bi-level cognitive orchestration framework that decouples network control at two scales: a strategic upper level uses a Large Language Model (LLM) with an offline-distilled Expert Knowledge Base (EKB) to dynamically infer preference weights based on real-time network conditions; a lower execution level employs a fidelity-aware genetic scheduler to compute collision-free joint routing and task offloading schedules. This approach tackles the challenge of spatio-temporal resource fragmentation caused by variations in onboard computing, energy availability, and intermittent connectivity.

Extensive evaluations on a high-fidelity Walker-Delta testbed under mixed-criticality workloads demonstrate significant improvements. The framework achieved a 30.7% reduction in packet loss, a 30% improvement in energy efficiency over the best learning-based baseline, and an 8.5% decrease in end-to-end latency. It also maintained robust performance under cascading node-failure scenarios. This work shows how combining RAG with LLMs can enable real-time, knowledge-guided optimization in distributed satellite edge computing systems.