Generative AI-Enabled Mission-Aware Radio Orchestration for RIS-Assisted LEO Satellite ISAC Systems

A mixed-timescale framework where an LLM translates operator mission language into structured policies, while deterministic physical-layer optimization enforces feasibility for integrated sensing and communication.

Fitsum Debebe Tilahun; Chung G. Kang· 2026· DOI 10.48550/arXiv.2608.25803

The core problem

Mission-adaptive low-Earth-orbit (LEO) satellite networks with integrated sensing and communication (ISAC) must retarget radio resources as operator goals change. Traditional radio resource management assumes fixed or slowly varying objectives, but emerging LEO ISAC missions require rapid reconfiguration across communication, sensing, and fairness priorities. The central challenge is bridging flexible, human-readable operator intent with the rigid feasibility constraints of the physical layer. To enable this adaptation from flexible operator language, the authors develop a generative-AI-enabled radio-orchestration framework. A large language model (LLM) maps each mission into a structured policy comprising communication, sensing, and fairness weights, mandatory quality-of-service (QoS) thresholds, power-allocation guidance, and solver initialization. Deterministic validation and physical-layer optimization then enforce feasibility and realize the policy through beam, power, and reconfigurable intelligent surface (RIS) configuration. This mixed-timescale design uses generative AI for semantic adaptation at the mission timescale while retaining conventional wireless optimization at

Innovation

The authors evaluate priority-order accuracy, defined as the fraction of missions where the LLM correctly orders the communication, sensing, and fairness priorities. On held-out instructions, LLM-ZS achieves accuracy, while LLM-ICL achieves . The improvement from ICL is mainly in numerical calibration, i.e., producing weights and thresholds closer to the optimal values.Downstream radio performance is measured after physical-layer optimization. Despite the accuracy difference, the downstream performance difference between LLM-ZS and LLM-ICL is statistically unresolved. The authors explain that both methods usually recover the hard constraints (e.g., , , ) that determine admissible actions. Once these constraints are satisfied, the exact weights have limited impact on the final radio performance.When active beams and RIS phases are optimized directly via explicit alternating optimization, the qualitative ordering of policies is preserved. This indicates that the LLM's policy structure is robust to the specific optimization algorithm used.Key quantitative findings:- LLM-ZS priority-order accuracy on held-out missions: .- L

Mission-adaptive low-Earth-orbit (LEO) satellite networks with integrated sensing and communication (ISAC) must retarget radio resources as operator goals change. Traditional radio resource management assumes fixed or slowly varying objectives, but emerging LEO ISAC missions require rapid reconfiguration across communication, sensing, and fairness priorities. The central challenge is bridging flexible, human-readable operator intent with the rigid feasibility constraints of the physical layer.

To enable this adaptation from flexible operator language, the authors develop a generative-AI-enabled radio-orchestration framework. A large language model (LLM) maps each mission into a structured policy comprising communication, sensing, and fairness weights, mandatory quality-of-service (QoS) thresholds, power-allocation guidance, and solver initialization. Deterministic validation and physical-layer optimization then enforce feasibility and realize the policy through beam, power, and reconfigurable intelligent surface (RIS) configuration. This mixed-timescale design uses generative AI for semantic adaptation at the mission timescale while retaining conventional wireless optimization at the faster channel timescale.

Why it matters

The results show how generative AI can enhance next-generation radio orchestration without replacing feasibility-critical physical-layer optimization. The mixed-timescale design is central: the LLM handles semantic adaptation at the mission timescale, where operator goals change, while conventional wireless optimization handles the faster channel timescale, where hard constraints must be enforced. The finding that LLM-ZS and LLM-ICL have statistically unresolved downstream performance differences suggests that for many missions, the hard constraints dominate the achievable performance. Thus, LLM-ZS is the low-context default, reducing prompt engineering effort and computational cost. LLM-ICL is useful for semantically difficult missions requiring finer calibration, such as those with unusual priority combinations or tight QoS thresholds. The explicit alternating optimization preserving qualitative ordering indicates that the framework is not overly sensitive to the specific solver. This is important for practical deployment, where different operators may use different physical-layer algorithms. Limitations and future work include: the evaluation is limited to specific mission types and channel conditions; the LLM's policy generation may fail on out-of-distribution missions; and the framework assumes a centralized orchestration entity. Future work could explore distributed orchestration, online learning of mission-policy mappings, and integration with other AI-native network functions. Overall, the paper demonstrates a viable path for mission-aware radio orchestration in RIS-assisted LEO satellite ISAC systems, where generative AI provides semantic flexibility and deterministic optimization ensures feasibility.

Who should read this

CS practitioners and researchers

Opening member content…