MimicAI: An LLM based system that mimics and explains

Large Language Models (LLMs) have demonstrated significant proficiency across various tasks, showcasing emergent abilities in in-context learning, instruction following, and multi-step reasoning. However, their operation as “black boxes” poses a challenge regarding explainability, affecting stakeholders’ trust in these models. In contrast, simpler AI models such as linear classifiers, decision trees, and random forests offer greater explainability. They are easier to train but typically require intricate tuning and do not generalize well across different objectives. These simpler models also demand extensive data engineering and machine learning expertise to optimize performance. To address these limitations, we introduce MimicAI, a novel system designed to replicate the functionality of simpler AI models while enhancing their interpretability and ease of use. MimicAI generates outputs that are easily understandable regarding input contributions and simplifies the model extension process for domain experts without the need for deep machine learning knowledge. By integrating the principles of in-context learning (ICL), MimicAI achieves superior generalizability without requiring extensive fine-tuning, making it a robust tool for applications seeking the dual benefits of simplicity and deep learning prowess.