LLM for AI Explainer

Explainable artificial intelligence (XAI) methods improve the interpretability of machine learning models, but existing techniques often present explanations in rigid, technical formats that limit accessibility for non-expert users. Large language models (LLMs) provide a promising way to generate flexible and contextually rich natural language explanations; however, challenges like hallucination and trustworthiness remain major concerns. This study uses LLMs to enhance explainability in Reinforcement Learning models by initially translating feature importance outputs into natural language explanations. We explore various LLM architectures, prompting strategies, and Retrieval Augmented Generation (RAG) techniques to produce explanations that are both flexible and reliable. We further develop a summarization technique to simplify the data sent to the LLM and propose an evaluation method to ensure the summarization provides adequate data coverage. Using this evaluation, we compare our system against existing methods. We also assess the trustworthiness of the explanations using consistency metrics. By bridging the gap between complex RL models and human interpretability, this research advances the deployment of Artificial Intelligence (AI) systems in real-world applications where transparency and trust are essential.