ENWAR 2.0: An Agentic Multimodal Wireless LLM Framework With Reasoning, Situation-Aware Explainability and Beam Tracking

The evolution of next-generation wireless networks demands intelligent, adaptive, and explainable decision-making for robust communication in dynamic environments. This paper presents <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Enwar 2.0</small>, the first agentic large language model (LLM) framework integrating adaptive retrieval-augmented generation (RAG) and chain-of-thought (CoT) reasoning into situation-aware and explainable wireless network management. <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Enwar 2.0</small> introduces two specialized agents: a transformer-fusion (TransFusion)-based beam prediction agent and an environment perception agent, both of which fuse multi-modal sensory inputs—including camera, LiDAR, radar, and GPS—from the DeepSense6G dataset. The beam prediction agent enables infrastructure-to-vehicle (I2V) target-in-the-loop beam tracking and real-time adaptation based on dynamic environmental conditions. In contrast, the environment perception agent provides situation-aware reasoning and justifications for beam decisions. Unlike its predecessor, <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Enwar 1.0</small>, which relied on static knowledge bases (KBs) and text-only LLMs, <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Enwar 2.0</small> is designed for CoT reasoning, leverages LLaMa3.2-3B/LLaMa3.1-8B/LLaMa3.3-70B for text-generation, the multi-modal capabilities of LLaMa 3.2, and employs LlamaIndex for fine-grained, dynamic context retrieval, eliminating retrieval ambiguities and enhancing response relevance. Numerical results show that the beam prediction agent achieves up to 90.0% Top-3 accuracy at <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$t+3$</tex-math></inline-formula>, effectively predicting optimal beam selections three time steps ahead. Overall, <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Enwar 2.0</small> achieves state-of-the-art performance, with up to 89.7%/83.5% interpretation/perception correctness, 81.6%/80.9% faithfulness, and 89.9%/88.2% relevancy. In comparison, the baseline pretrained LLaMa3 models without adaptive RAG achieves up to 80.3%/77.3% correctness, and the baseline without RAG performs significantly worse at 67.1%/64.8%. Additionally, <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Enwar 2.0</small> reduces processing time by over 100% relative to the baseline, while its adaptive RAG improves performance by up to 13.7% compared to static RAG.