STR-Agent: An LLM-Driven Agent for QoS-Aware Routing in LEO Satellite Networks

Summary

The paper proposes STR-Agent, an LLM-driven agent for QoS-aware routing in LEO satellite networks, targeting the gap between natural-language service requests and adaptive routing decisions. Four modules form a closed loop: a Perception module (LoRA fine-tuned Qwen2.5-7B) parses requests into source/destination coordinates and one of three service types (delay-sensitive, bandwidth-sensitive, best-effort); an Execution module routes hop-by-hop through tool calls to three Dijkstra variants with service-specific edge weights; an Experience Buffer aggregates hop-level outcomes; and a Reflection module re-maps service types to routing tools under severe congestion using hand-crafted thresholds. The SFT dataset holds 33,000 samples generated by Qwen2.5-7B itself, using 3GPP TR 22.822 scenarios and 49 cities. In a 1,584-satellite Walker-Delta simulation, intent accuracy rises from 45.4% (base) to 92.45% (SFT); at 270 Mbps, STR-Agent (SFT) reports 260.11 ms delay-sensitive delay versus 696.01 ms for QSMR (the abstract claims up to 60% reduction versus DQ-Dijkstra), with hop counts dropping from 53.03 to 25.94 and better queue metrics. The reflection ablation saves 120 ms at 600 Mbps. Main limitations: training and test data both come from the same LLM generator with no human-labeled evaluation; baselines exclude learning-based methods and no rule-based-dispatcher control isolates the LLM's contribution; no variance or statistical testing; single 10-second simulation runs; and per-hop LLM inference overhead is unmeasured.