Technical Summary: ENWAR 3.0

Problem Statement

Maintaining robust millimeter-wave (mmWave) connectivity in vehicular networks requires real-time adaptation to environmental dynamics, sensor degradation, and link variability. Traditional network control mechanisms often rely on rule-based heuristics or static beam tracking, which fail to scale in multi-modal sensing environments. These conventional approaches cannot effectively adapt decisions based on sensor reliability, environmental context, or historical system behavior. Furthermore, existing Large Language Model (LLM) solutions for wireless control often lack real-time decision-making capabilities, remain text-centric, or fail to integrate multi-modal sensory data (LiDAR, radar, cameras, GPS) with transparent, explainable logic. There is a critical gap in designing seamless frameworks that blend multi-modal sensing, interpretable reasoning, and adaptive decision-making under strict real-time constraints for Infrastructure-to-Vehicle (I2V) communications.

Methodology

ENWAR 3.0 is an environment-aware reasoning framework that unifies multi-modal sensing, agentic LLMs, and context-driven model selection. The system operates within a 300ms control interval, managing beamforming, blockage prediction, and handover management through a hierarchical control framework.

Core Architecture

1. Multi-Modal Perception and Degradation Detection:

   • The system ingests synchronized data from cameras, LiDAR, radar, and GPS.
   • A dedicated Environment Classifier (a multi-output Random Forest) detects real-time sensor degradation (e.g., fog, occlusion, noise) with over 99% accuracy. It outputs binary flags to exclude unreliable modalities dynamically.
   • A synthetic degradation pipeline was developed to train this classifier, simulating adverse conditions like blurring, sparsification, and signal jamming.

2. Agentic Orchestration via Primed LLM:

   • A centralized LLM acts as the orchestrator, primed via Chain-of-Thought (CoT) prompting and human-in-the-loop (HITL) feedback.
   • The LLM interprets structured environment states and sensor health flags to invoke specialized agents (Beam Prediction, Blockage Prediction, Handover, Environment Perception) exactly once per cycle.
   • The system employs Long-Term Memory (using LlamaIndex) to persist historical patterns, such as blockage durations and past agent decisions, enabling temporal continuity in reasoning.

3. Adaptive Model Routing (DRL):

   • A Deep Reinforcement Learning (DRL) agent, trained with Proximal Policy Optimization (PPO), dynamically selects the optimal combination of sensor modalities for prediction tasks.
   • The system maintains a repository of 15 pretrained model variants per agent (covering various sensor combinations). The DRL policy routes tasks to the most suitable variant based on current sensor health and historical context, avoiding degraded sensors.

4. Task-Specific Agents:

   • Beam Prediction: Anticipates optimal beam indices using multi-modal inputs to avoid exhaustive beam sweeping.
   • Blockage Prediction: Estimates the likelihood of occlusion events up to 5 time steps (1.5s) into the future using late fusion of modality-specific predictions.
   • Handover Management: Triggers handovers to secondary base stations (BS2) based on a dual-trigger policy: significant power superiority or persistent blockage (exceeding 5 time steps) combined with power improvement.

5. Response Delivery:

   • The system generates structured, interpretable responses that integrate perception, prediction, and control decisions.
   • Critical control-path decisions (agent invocation, beam selection) are bounded within the 300ms sampling window, while detailed natural language justifications are generated asynchronously to ensure real-time performance.

Key Contributions

The paper outlines five primary contributions:

• Hierarchical Agent Orchestration: Reformulating multi-modal wireless control as a hierarchical agent problem, utilizing a modular LLM-driven architecture with CoT priming and RLHF to coordinate beam/blockage prediction and handover decisions.
• Degradation-Aware Control Layer: Introducing a synthetic sensor degradation pipeline and a multi-output classifier (99.1% accuracy) that enables dynamic exclusion of unreliable modalities, replacing static fusion strategies with reliability-aware sensing.
• DRL-Based Adaptive Model Routing: Designing a degradation-aware DRL policy that dynamically selects among 15 pretrained modality combinations per agent, adapting to sensor status rather than relying on fixed model invocation.
• Persistence-Aware Blockage and Handover Control: Integrating blockage duration tracking and long-term memory into a policy-triggered handover mechanism, creating a closed-loop beam–blockage–handover pipeline absent in prior iterations.
• Real-Time, Bounded-Latency Operation: Demonstrating that the full system maintains a worst-case end-to-end control-path latency of 289.7ms within a 300ms sampling interval, ensuring compatibility with real-time I2V requirements.

Results

Extensive evaluations were conducted across 15 sensor combinations and various LLM architectures (ranging from 3B to 70B parameters):

• Predictive Performance:
  • Beam Selection: Achieved up to 88.5% Top-3 accuracy (with Camera+GPS+LiDAR) and an average power loss of -0.009 dB.
  • Blockage Prediction: Achieved 98.4% F1-score (with Camera+Radar) and an AUC-ROC of 0.988.
  • Environment Classification: Detected sensor degradation with 99.1% accuracy across modalities.
• Reasoning Correctness:
  • The system achieved 87.3% reasoning correctness with a 3B parameter model, improving to 89.3% with a 70B model. This represents a significant improvement over vanilla baselines (69.4%).
  • Ablation studies showed that priming (82.5%) and CoT reasoning (82.2%) provided the largest individual gains, while Long-Term Memory and DRL routing further enhanced stability.
• Latency and Efficiency:
  • The worst-case end-to-end latency for the control path was 289.7ms, fitting within the 300ms sampling window.
  • Larger models (70B) offered only marginal performance gains (approx. 2% improvement) but incurred significantly higher inference latency (3.08s vs. 2.17s for 3B), suggesting diminishing returns for real-time deployment.
• Robustness:
  • Under edge cases (e.g., heavy fog, GPS jamming, LiDAR failure), the system maintained high performance by dynamically switching to fallback models (e.g., radar-only or camera-radar fusion), demonstrating graceful degradation.
  • The DRL agent selected optimal actions in 86.6% of episodes, significantly outperforming rule-based (67.1%) and random (13.2%) baselines.

Significance

The paper positions ENWAR 3.0 as a scalable foundation for LLM-integrated wireless systems capable of reasoning, perceiving, and adapting in real-time. It advances the field from perception-assisted decision support to policy-driven, degradation-aware orchestration. By establishing a unified real-time loop for beam prediction, blockage detection, and handover control, the work addresses the limitations of static heuristics and black-box reasoning in dynamic vehicular environments. The framework demonstrates that integrating structured reasoning, memory, and adaptive routing allows wireless systems to maintain high performance and interpretability even under significant sensor uncertainty and environmental dynamics.