PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Fitsum Debebe Tilahun, Chung G. Kang

arXiv:2608.25803v1cs.NIeess.SY

TL;DR

Mission-adaptive LEO satellite ISAC must retarget radio resources as operator priorities change. The paper compiles natural-language missions into structured policies and uses deterministic optimization for feasible beam, power, and RIS control, achieving strong held-out priority recovery while leaving downstream zero-shot versus in-context differences statistically unresolved.

  • Problem

    Mission changes require different radio configurations, but operator intent is expressed in language rather than optimization variables.

  • Method

    An LLM maps missions to objective weights, hard QoS thresholds, power guidance, and solver initialization; deterministic validation and physical-layer optimization execute the policy.

  • Results

    LLM-ZS and LLM-ICL achieve 91.7% and 94.4% priority-order accuracy, respectively, on held-out missions, while their downstream performance difference is statistically unresolved.

  • Takeaways & Limitations

    LLM-ZS is the low-context default, whereas LLM-ICL is useful for semantically difficult missions requiring finer calibration.

Abstract

from arXiv · show

Mission-adaptive low-Earth-orbit (LEO) satellite networks with integrated sensing and communication (ISAC) must retarget radio resources as operator goals change. To enable this adaptation from flexible operator language, we develop a generative-AI-enabled radio-orchestration framework in which a large language model (LLM) maps each mission into a structured policy comprising communication, sensing, and fairness weights, mandatory quality-of-service 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. We compare zero-shot (LLM-ZS) and in-context (LLM-ICL) operation on familiar and held-out compositional missions. On held-out instructions, LLM-ZS and LLM-ICL achieve $91.7\%$ and $94.4\%$ priority-order accuracy, respectively, with ICL mainly improving numerical calibration. Their downstream radio-performance difference is statistically unresolved because both usually recover the hard constraints that determine admissible actions. Accordingly, LLM-ZS is the low-context default, while LLM-ICL is useful for semantically difficult missions requiring finer calibration. Explicit alternating optimization preserves the qualitative ordering when active beams and RIS phases are optimized directly. The results show how generative AI can enhance next-generation radio orchestration without replacing feasibility-critical physical-layer optimization.

I. INTRODUCTION

LEO satellite ISAC combines difficult, time-varying wireless conditions with missions that may prioritize throughput, sensing, or fairness differently. The framework uses an LLM to compile language into structured priorities while deterministic optimization realizes feasible beam, power, and RIS configurations.

  • LEO operation faces large propagation loss, changing geometry, and frequent beam-reconfiguration demands.
  • Different missions require different allocations of transmit power and RIS-assisted gain, even under the same channel realization.
  • The LLM generates continuous objective weights, mandatory QoS thresholds, and low-dimensional search guidance for unseen communication, sensing, and fairness combinations.
  • The architecture separates semantic adaptation from fast channel control through prompting modes, validation, caching, and solver-based execution.
  • The physical model includes satellite-to-RIS, direct, and RIS-assisted user channels, with communication and sensing streams jointly transmitted.

B. Sensing Model

The sensing model represents target illumination through direct and RIS-assisted paths and evaluates sensing quality using target-return power, sensing SNR, and angle-estimation accuracy. Communication and sensing remain coupled through shared transmission resources.

  • The target channel combines direct satellite-to-target and RIS-assisted propagation paths.
  • The monostatic radar relation models nominal target-return power using transmit power, antenna gains, radar cross section, wavelength, range, and additional losses.
  • Both communication beams and the dedicated sensing beam illuminate the target, coupling sensing SNR to the shared transmit design.
  • The angle-estimation CRB depends on sensing SNR, snapshot count, and the array’s angular sensitivity.
  • The model considers one dominant target and neglects clutter.

C. Mission-Aware Policy and Optimization Problem

The mission policy maps natural-language intent into weighted communication, sensing, and fairness utilities, hard QoS thresholds, and solver guidance. A deterministic non-convex optimizer then selects physical controls subject to those requirements and RIS constraints.

  • The policy vector contains normalized communication, sensing, and fairness weights, minimum rate and sensing-SNR thresholds, a power fraction, and solver initialization.
  • Bounded utilities normalize aggregate rate, sensing quality, and weakest-user fairness before combining them.
  • The mission-weighted utility rewards the three objectives while penalizing energy consumption and RIS reconfiguration.
  • The physical-layer controller selects beamformers, power split, and RIS coefficients by solving the policy-weighted optimization problem.
  • The optimization enforces user-rate thresholds, transmit-power allocation constraints, and unit-modulus RIS phases despite non-convex coupling.

A. Mixed-Timescale Architecture

The mixed-timescale architecture invokes the LLM when mission intent changes, then validates and caches its policy for faster channel-timescale optimization. Zero-shot and in-context modes differ by whether mission-policy demonstrations are supplied.

  • A. Mixed-Timescale Architecture: At the slower mission timescale, the LLM converts operator instructions into priorities, requirements, and solver guidance, which a validator checks and caches.
  • A. Mixed-Timescale Architecture: At the faster channel timescale, deterministic optimization maps the policy and channel state to feasible beams, power allocation, and RIS phases.
  • A. Mixed-Timescale Architecture: The LLM is invoked again only when the mission changes, while the radio solver determines how priorities are physically realized.
  • A. Mixed-Timescale Architecture: The prompt can include an optional set of example mission-policy pairs called demonstrations.
  • A. Mixed-Timescale Architecture: The LLM maps mission language to weights, QoS thresholds, power guidance, and initialization choices, producing an unvalidated policy whose thresholds determine admissibility.

B. Zero-Shot and In-Context Learning Modes

The framework evaluates the same frozen LLM in zero-shot and in-context modes, differing only in whether example mission-policy pairs are supplied.

  • LLM-ICL differs by including example mission-policy pairs in the prompt.The two prompting modes use the same frozen LLM; only the inclusion of demonstrations changes.
  • LLM-ZS uses no demonstrations and relies on the rubric, output schema, and policy-field definitions for each new mission instruction.Multiple responses are generated and aggregated, but no demonstration is supplied and no model parameter is updated.

2) In-context learning mode (LLM-ICL):

In-context learning supplies leakage-controlled mission-policy demonstrations that illustrate how priorities, requirements, and negation map into structured radio policies. The resulting proposals are aggregated, deterministically validated, and executed through a shared radio evaluator and solver.

  • In-context learning mode (LLM-ICL):: LLM-ICL adds D example instruction-policy pairs to show representations of ordered priorities, co-primary objectives, and comparisons.The demonstrations affect only the current response; the model is not fine-tuned.
  • In-context learning mode (LLM-ICL):: A communication objective that is secondary but mandatory is represented by a smaller communication weight and a nonzero rate threshold.
  • In-context learning mode (LLM-ICL):: The two prompting modes share the model, schema, response aggregation, validator, and physical solver, isolating contextual examples as the comparison variable.Demonstrations, QoS pairs, and wording are disjoint from the held-out evaluation set.
  • In-context learning mode (LLM-ICL):: Multiple structured proposals are consolidated using componentwise medians for numerical fields and majority votes for categorical guidance.
  • In-context learning mode (LLM-ICL):: Validation parses the schema, clips values, normalizes priorities, verifies QoS thresholds, and maps initialization to supported solver modes.Only the validated policy is cached and passed to the radio solver.
  • In-context learning mode (LLM-ICL):: The shared evaluator uses a common candidate action set while reporting communication, sensing, fairness, energy, and reconfiguration metrics.This shared action space separates mission interpretation from deterministic radio execution.
  • In-context learning mode (LLM-ICL):: Compiled thresholds filter admissible configurations, mission weights rank them, and initialization guidance steers the non-convex search.If no candidate satisfies all requirements, the solver first minimizes normalized QoS violation.

IV. NUMERICAL EVALUATION

The numerical evaluation studies a single RIS-assisted LEO satellite ISAC setting and uses two complementary language-benchmark instruction sets. The setup varies RIS size while testing known mission types.

  • The simulation uses one 8-element LEO satellite array, three receiver groups, one target, and a fixed ground RIS.The satellite altitude, carrier frequency, and bandwidth are 600 km, 20 GHz, and 20 MHz.
  • The primary RIS configuration has N = 32 elements, while N ∈ {8, 16, 32, 64} is used to examine surface-size effects.
  • The benchmark contains 36 mission prompts for evaluating mission-policy interpretation.
  • The language benchmark contains two complementary instruction sets: familiar mission types and held-out combinations outside the established taxonomy and demonstrations.

2) Held-out instructions:

Held-out evaluation tests compositional instructions involving priorities, comparisons, and negation, using leakage-controlled paired prompts and a common radio simulator. Performance is assessed against an oracle reference under the intended mission policy.

  • Held-out instructions:: Held-out instructions combine priorities, comparisons, and negation absent from both the classifier taxonomy and in-context demonstrations.
  • Held-out instructions:: The LLM compiler is queried three times per mission, then aggregates numerical fields by componentwise median and initialization mode by majority vote.DeepSeek-V4-Flash is used with sampling temperature 0.05.
  • Held-out instructions:: LLM-ZS receives no demonstrations, whereas LLM-ICL receives three leakage-controlled mission-policy examples.Prompts, compiled policies, channel realizations, and the Monte Carlo seed are frozen for paired evaluation.
  • Held-out instructions:: Every method selects from the same precomputed communication-power fractions and RIS operating modes evaluated by a common simulator.The selected beam, power, and RIS configuration is evaluated using the intended mission weights and QoS thresholds.
  • Held-out instructions:: Normalized regret compares each method’s selected candidate with the best candidate under the intended mission policy.A smaller R_m indicates less loss in mission-weighted radio utility.
  • Held-out instructions:: Confidence intervals use a three-level bootstrap over mission families, wording variants, and channel realizations.
  • B. Performance Analysis: The proposed compiler is compared with language, policy, and physical-layer baselines to isolate semantic adaptation, validation, and radio-execution effects.An oracle reference uses the stored mission specification to select the best candidate and define the reference.

1) TF-IDF classifier:

The study compares mission-policy baselines and LLM modes, showing that generative compilation and explicit radio optimization preserve mission-aware performance across unfamiliar instructions and solver settings.

  • Physical-layer validation: At N = 32, optimized RIS phases make joint communication and sensing requirements approximately six times more likely to be satisfied than random phases.Phase optimization also lowers the angle-estimation CRB by about 85% at the same surface size.
  • Mission-policy compilation: LLM-ZS is the lower-context default, while LLM-ICL mainly helps linguistically difficult missions requiring finer numerical calibration.On held-out missions, both modes remain above 90% on compositional and negated instructions, while classifier and retrieval approaches degrade more sharply.
  • Mission-policy compilation: The closed-set classifier maps instructions to familiar classes and cannot represent unseen combinations, whereas keyword rules and demo retrieval cannot synthesize new priority combinations.The classifier's held-out weight error is 0.3676, compared with 0.1607 for LLM-ICL.
  • Radio-action consequences: Both LLM modes remain closest to the oracle on held-out missions, while fixed and retrieval-based policies incur substantially larger mission-weighted utility loss.The evaluator compares configurations after feasibility filtering, utility ranking, and search guidance.
  • Radio-action consequences: LLM-ICL's mean regret is 0.0060 versus 0.0089 for LLM-ZS, but their paired difference confidence interval includes zero.Both modes usually recover mandatory thresholds, so weight differences often select the same or nearly equivalent radio configuration.
  • Solver-level validation: Under explicit alternating optimization of beams, power split, and RIS phases, the two LLM modes remain the leading mission-aware methods.Random RIS phases incur more than two orders of magnitude greater regret.

V. CONCLUSION AND FUTURE WORK

The framework translates natural-language missions into validated radio policies while deterministic optimization handles feasibility-critical wireless control. Both prompting modes generalize beyond fixed policy taxonomies, with zero-shot as the default and in-context learning reserved for semantically difficult missions.

  • The LLM converts natural-language missions into validated utility priorities, hard QoS requirements, and solver guidance for RIS-assisted LEO satellite ISAC.
  • LLM-ICL provides finer semantic calibration, but its lower mean regret of 0.0060 versus 0.0089 for LLM-ZS is not statistically resolved.
  • Both prompting modes generalize to compositional and negated missions beyond a fixed policy taxonomy.
  • LLM-ZS is the low-context default, whereas LLM-ICL is most useful when semantically difficult missions justify additional context.
  • Explicit alternating optimization preserves the qualitative ordering when active beams and RIS phases are optimized directly.
Loading 2608.25803v1…