Explaining black-box behavior in large language models

Abstract

Large language models (LLMs), such as GPT and BERT, have transformed natural language processing (NLP), achieving state-of-the-art performance across diverse tasks. However, their opaque decision-making processes raise critical concerns about transparency, trust, and accountability, particularly in high-stakes domains. This paper reviews key interpretability techniques, including attention visualization, layer wise relevance propagation, feature attribution, and counterfactual analysis, and proposes a comprehensive taxonomy for explaining LLM behavior. Additionally, we address the ethical implications of explainability, such as bias mitigation, privacy preservation, and regulatory compliance. By synthesizing current methods and identifying challenges, this work lays the foundation for future research into interpretable and ethically responsible LLMs.