- The paper demonstrates a formally verified pipeline that translates natural language intents into enforceable routing constraints using LLMs and formal validation.
- It employs a GNN-based cost-to-go router that achieves 17× speedup over Dijkstra while ensuring near-perfect packet delivery even under severe network constraints.
- The approach integrates a deterministic multi-pass validator to robustly detect infeasible or adversarial constraints, guaranteeing zero unsafe acceptances.
Validated Intent Compilation for Constrained Routing in LEO Mega-Constellations
Problem Motivation and Contributions
The deployment of large-scale Low Earth Orbit (LEO) satellite constellations introduces a complex, dynamic network environment with unique routing challenges. The critical requirement is translating high-level operator intents—articulated in natural language—into enforceable, low-level routing constraints on rapidly evolving topologies characterized by periodic polar link dropouts, frequent node or link failures, and stringent performance goals. Manual translation, common in current practice, is infeasible due to the scale, dynamism, and safety-critical nature of these networks.
This paper presents a formally verified end-to-end pipeline for intent-driven, constrained routing in LEO mega-constellations, with three primary contributions:
- A cost-to-go router using graph neural networks (GNNs), providing orders-of-magnitude speedup over traditional Dijkstra algorithms while matching routing quality.
- A LLM-based intent compiler that robustly maps operator natural language intents to a typed intermediate representation (ConstraintProgram) of routing constraints, significantly advancing over rule-based approaches.
- A deterministic, multi-pass constraint validator that certifies both structural and semantic correctness, achieving full safety with zero unsafe constraint acceptances and comprehensive detection of infeasibility or adversarial errors.
System Architecture and Methodology
Central to the pipeline is the ConstraintProgram IR, a typed, formal schema capturing the semantics of operator intents. It supports diverse hard and soft constraint types—covering node/plane/edge disablement, region/latitude avoidance, flow deadlines, hop limits, k-edge-disjointness, and minimum capacity reserves—enabling compositional representation amenable to downstream validation and routing.
Intent Compilation with LLMs
To bridge the semantic gap between human expressions and formal constraint schemas, the system employs a Qwen3.5-9B LLM with few-shot prompting. The prompt articulates detailed schema definitions, permissible values, and multiple intent examples, enabling the model to decompose and translate compositional and conditional intents. Crucially, the compilation operates in a closed-loop with the validator: verifier error messages prompt repair iterations, which empirically recover an additional 20.5% from initial failures, yielding a 98.4% overall compilation rate and 87.6% full semantic match on feasible intents.
The LLM-based approach shows a strong performance advantage over both smaller (4B) models and rule-based baselines, with a +46.2 percentage point full semantic match gain over rule-based parsing for compositional intents. Out-of-distribution (OOD) evaluation demonstrates robust generalization, with only minimal degradation in accuracy for paraphrased or ambiguous intents.
Deterministic Multi-Pass Constraint Validation
The validator comprises eight sequential passes:
- Schema validation
- Entity grounding against the constellation model
- Type safety between constraints and target entities
- Value range enforcement
- Conflict and contradiction detection
- Physical admissibility (e.g., SLOs below physical minima)
- Topological reachability
- Feasibility certification via constructive routing witnesses (BFS, Dijkstra, Edmonds-Karp, etc.)
Certification is strictly constructive: only intents for which a routing witness exists are accepted; unsupported or infeasible combinations receive hard rejection or abstention (deferred to fallback Dijkstra routing), resulting in zero occurrences of unsafe acceptance in all evaluated infeasible cases, including adversarial settings. Validator runtime is sub-millisecond, preserving real-time applicability.
GNN-based Cost-to-Go Routing
The router is a 3-layer graph attention network with supervised distillation from Dijkstra next-hop policy. Node-level features encode orbital and topology-specific information, ensuring generalization over sampled constellation states. The GNN routes on masked topologies reflecting active constraint sets without retraining. The architectural speedup (17×) enables feasible per-packet real-time routing at constellation scale.
Experimental Validation
Benchmark Construction and Metrics
A synthetic but rigorous benchmark of 240 natural language intents (single, compositional, conditional, and infeasible) was constructed with matched ground-truth constraint programs. Metrics detail compilation, semantic match, packet delivery ratio (PDR), violation rate, OOD generalization, and adversarial robustness.
GNN Routing Quality
The GNN achieves near-perfect routing quality—99.8% PDR overall, matching Dijkstra within 0.22pp even under plane failures or aggressive topology degradation (up to 85% capacity removal). Critically, a reachability-conditioning analysis shows that apparent drops in raw PDR under severe constraints (e.g., polar avoidance) are fully explained by topological reachability limits, not routing algorithm limitations; both GNN and Dijkstra achieve 100% delivery on reachable pairs.
The LLM compiler demonstrates strong semantic match (87.6%) and compositionality (86.2% on intents with multiple constraints). Ablation studies highlight the dominant role of few-shot prompting and the verifier repair loop. The 9B model significantly outperforms both 4B and rule-based systems. Latency is compatible with operator workflows involving minute-level intent compilation; further gains are possible via model distillation.
Validator Safety and Adversarial Resilience
The validator achieves 100% detection of all infeasible and adversarially crafted constraints. There are no instances of unsafe program acceptance. All edge-case detection—including contradictory, structurally invalid, or physically infeasible intents—are handled with definitive failure or abstention, preserving system safety guarantees.
Robustness and Generalization
OOD generalization is robust for single and conditional intents, with some degradation in complex compositional phrasing. The GNN's topology generalization is limited: architectures generalize across altitude variation but not orbital inclination or ISL layout, confirming that mask-based constraint enforcement and formal validation are necessary for architecture independence.
Implications and Future Directions
The system demonstrates that the semantic intent-to-constraint compilation pipeline, backed by deterministic validation and topology-masked GNN acceleration, enables intent-based networking at the scale and safety requirements demanded by LEO mega-constellations. The results confirm the practical viability of LLM-based semantic parsing in high-stakes network infrastructure, provided that model flexibility is contained by formal, deterministic safety checks.
Theoretically, the work formalizes a methodology for closing the semantic gap with constructive certification, which can generalize to other domains requiring intent-driven control with hard safety constraints. Practically, this architecture is directly transferable to production satellite network control planes, and extensible to heterogeneous network domains.
Future research avenues include:
- Constraint Pre-solving and Satisfiability: Integrating pre-solvers could further close semantic gaps for infeasible compositions unhandled by current routing fragments.
- Constraint-aware GNN Routing: Conditioning GNN inference directly on constraint features may improve OOD routing generalization and constraint-specific optimization.
- Real-world Operator Traces: Validating the entire pipeline on annotated production intent logs is essential to confirm robustness against natural language ambiguity and operational diversity.
- Constellation-agnostic Learning: Developing GNN/routing architectures that are fundamentally invariant across constellation geometries to eliminate retraining requirements.
Conclusion
This work establishes a rigorously validated, safe, and efficient pipeline for translating operator intents into enforceable routing policies in LEO mega-constellations. By integrating LLM-based semantic parsing, formal constraint representation and certification, and GNN routing, the approach guarantees both expressiveness and safety. The experimental validation substantiates all primary claims, demonstrating significant quantitative advances in intent-driven network management. The framework is positioned as a reference methodology for future AI-driven control architectures in evolving, safety-critical network domains.