Validated Intent Compilation for Constrained Routing in LEO Mega-Constellations
Summary: arXiv:2604.07264v1 Announce Type: cross
Abstract: Operating Low Earth Orbit (LEO) mega-constellations requires translating high-level operator intents into low-level routing constraints. This task involves both natural language understanding and network-domain expertise. In this article, we present an end-to-end system that comprises three innovative components designed to enhance routing efficiency and effectiveness.
Components of the System
-
GNN Cost-to-Go Router
This component distills Dijkstra-quality routing into a graph attention network with 152,000 parameters. It achieves a remarkable 99.8% packet delivery ratio while providing a 17x inference speedup compared to traditional methods.
-
LLM Intent Compiler
The LLM intent compiler converts natural language intents into a typed constraint intermediate representation. Utilizing few-shot prompting along with a verifier-feedback repair loop, it has achieved a 98.4% compilation rate and an 87.6% full semantic match on feasible intents across a 240-intent benchmark, which includes 193 feasible and 47 infeasible intents.
-
Deterministic Validator
This 8-pass validator focuses on constructive feasibility certification, achieving 0% unsafe acceptance for all 47 infeasible intents. It successfully detects 100% of structural corruption across 240 tests and maintains a flawless record against 15 targeted adversarial attacks.
End-to-End Evaluation
The system underwent comprehensive end-to-end evaluation across four constrained routing scenarios. Impressively, it confirmed zero constraint violations with both routers employed in the system. Furthermore, analyses in polar-avoidance scenarios revealed that apparent performance gaps are primarily attributed to topological reachability ceilings rather than the routing quality itself.
Performance Comparison
Notably, the LLM compiler outperformed a traditional rule-based baseline by an impressive 46.2 percentage points when handling compositional intents. This significant improvement underscores the efficacy of leveraging advanced language models in routing configuration processes.
Conclusion
In summary, this innovative system effectively bridges the semantic gap between operator intent and network configuration. By maintaining the necessary safety guarantees for operational deployment, it represents a critical advancement in the management of LEO mega-constellations. As the demand for efficient and reliable network management continues to grow, solutions like this will be pivotal in ensuring optimal performance and safety in dynamic routing environments.