Wireless Knowledge Grounding in Smaller Llms Using Retrieval Augmented Generation and Fine-Tuning

While LLMs have shown tremendous progress in broad natural language processing (NLP) tasks, their application to solving domain-specific NLP problems (e.g., wireless) has been understudied. Existing research on LLMs for the wireless domain is mostly limited to using retrieval-augmented generation or model fine-tuning for wireless question-and-answering (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathrm{Q} / \mathrm{A}$</tex>) tasks using bigger models such as GPT, Mixtral, and Llama. In contrast to state-of-the-art models, in this paper, a smaller LLM such as Phi2 is evaluated on wireless Q/A tasks. A combination of RAG, model fine-tuning and hyper-parameter optimization is proposed to improve the accuracy of Phi 2 for wireless Q/A tasks. Experimental results demonstrate the proposed method's superior performance compared to vanilla LLMs and state-of-the-art LLMs with either fine-tuning or RAG. Comparing the results to larger models like Llama and Phi3 suggests that transitioning towards bigger models and relying on their extensive knowledge retrieval capabilities may not be the best approach for deploying LLMs on edge devices. This is because edge devices have limited computing and storage capabilities and are expected to operate in real-time for future wireless applications. Instead, the focus should be on improving the grounding of LLMs in wireless concepts using smaller models such as Phi2.