LLM-Based Digital Twin Intelligence for Application-Aware Network Selection in 6G Heterogeneous Wireless Networks

Technical Summary: LLM-Based Digital Twin Intelligence for Application-Aware Network Selection in 6G Heterogeneous Wireless Networks

Problem Statement
Future 6G heterogeneous wireless networks (HWNs) must support diverse radio access technologies (RATs) and applications with varying quality-of-service (QoS) requirements. Traditional network selection (NS) mechanisms, which rely on instantaneous radio measurements and static ranking rules (e.g., Multi-Attribute Decision Making or MADM), face significant limitations. These include the rank-reversal problem (RRP), where rankings change inconsistently when candidates are added or removed; an inability to autonomously translate high-level service intent into specific decision weights; and a tendency to trigger unnecessary handovers (HOs) due to sensitivity to instantaneous fluctuations. Existing approaches often fail to jointly capture the evolving wireless state, packet-level QoS behavior, and historical decision context required for stable, application-aware selection.

Methodology
The paper proposes a framework that shifts NS from an instantaneous decision-matrix operation to a decision process conditioned on an evolving, cross-layer Digital Twin (DT) state. The framework consists of four layers:

1. Physical-to-DT Construction: A site-specific 3D scene is generated using OpenStreetMap and Blender. Geometry-aware channel descriptors (channel gain, delay spread, path count) are extracted using Sionna ray-tracing (RT) based on the user equipment (UE) position and RAT deployment.
2. Cross-Layer Synchronization: The geometry-aware descriptors are mapped to packet-level QoS indicators (delay, jitter, packet loss ratio, energy efficiency, and access cost) using ns-3 packet-level emulation. This creates a unified decision matrix reflecting the actual service-level impact of each candidate RAT.
3. LLM-Based DT Intelligence: A Unified Intent Agent (UIA) translates user and service requirements into structured decision priorities. The framework employs two complementary NS branches:
   • MADM–LLM–NS: The LLM derives service-aware criterion weights, which are fed into a deterministic MADM operator (SAW or TOPSIS). To mitigate RRP, this branch utilizes History-Aware Adaptive Normalization (HAAN), which derives normalization references from a DT memory of past decisions rather than the instantaneous candidate set alone.
   • LLM–NS: The LLM performs direct listwise ranking of candidates. This branch employs Retrieval-Augmented In-Context Learning (RA–ICL) constrained by the Independence of Irrelevant Alternatives (IIA) principle. It retrieves similar historical DT states to guide ranking, preserving the relative order of retained candidates when RATs are withdrawn (via projection) while allowing adaptation when new RATs are inserted.
4. Evaluation and Memory Update: The selected RAT is evaluated against QoS targets, and the decision context and outcome are stored in the DT memory for future reuse, enabling the system to learn from past stability and performance.

Key Contributions

• DT-Grounded Formulation: The authors formulate NS as a decision task over an evolving cross-layer DT state that integrates geometry-aware channel descriptors, packet-level QoS, service intent, and decision memory, rather than relying on isolated instantaneous matrices.
• Unified Intent Agent (UIA): A mechanism that translates high-level service/user intent into structured, machine-readable decision priorities (weights and importance labels) for both MADM and direct LLM ranking branches.
• Stabilization Mechanisms: The introduction of HAAN to reduce candidate-set sensitivity in MADM ranking and RA–ICL to stabilize LLM-based ranking under candidate-set evolution, specifically addressing the rank-reversal problem.
• Comprehensive Evaluation: The framework is evaluated against representative MADM baselines (TOPSIS-AHP, SAW-AHP) across five 3GPP-aligned service profiles, analyzing RRP robustness, QoS satisfaction, unnecessary VHO rates, and computational complexity.

Numerical Results

• Rank Reversal Robustness: The proposed methods significantly reduce RRP compared to AHP baselines. SAW–LLM–NS achieved 0% RRP across all candidate-set perturbations. LLM–NS showed 0% RRP for candidate withdrawals (due to structural projection) and maintained low RRP (18.5%–41.0%) for insertions, reflecting a deliberate stability-adaptivity trade-off.
• QoS Satisfaction: For the demanding VR/AR service profile, LLM-based schemes improved reliability-oriented selection. LLM–NS achieved the lowest Packet Loss Ratio (0.26%) compared to 0.41% for TOPSIS-AHP, while maintaining competitive delay and jitter.
• Handover Stability: The MADM–LLM–NS branches significantly reduced unnecessary VHO rates. SAW–LLM–NS reduced unnecessary handovers by 47.9% compared to its SAW-AHP baseline, and TOPSIS–LLM–NS achieved a 28.1% reduction compared to TOPSIS-AHP.
• Complexity: The framework separates scenario-level DT generation costs from per-epoch decision costs. The LLM–NS branch is computationally efficient when memory hits occur (reusing past rankings), though it incurs higher costs on memory misses requiring fresh listwise ranking.

Significance and Claims
The paper claims that the novelty lies not merely in combining DTs, MADM, and LLMs, but in reformulating NS as an LLM-assisted, DT-grounded decision process. The authors assert that by operating over an evolving DT state and utilizing decision memory, the framework overcomes the limitations of static matrix-based approaches. Specifically, the work demonstrates that:

1. Service intent can be autonomously translated into adaptive selection behavior without manual weight configuration.
2. Decision stability under candidate-set evolution can be achieved through memory-driven normalization and retrieval-augmented reasoning.
3. The proposed framework offers a practical path toward intent-driven networking in 6G, balancing service-aware QoS satisfaction with mobility stability, without acting as a physical-layer predictor but rather as a decision-layer intelligence.

The authors position this work as a step toward autonomous, application-aware network management, distinct from prior studies that use LLMs merely as interfaces or DTs solely for simulation.