Measuring the Interpretability and Explainability of Model Decisions of Five Large Language Models
Kaito Fujiwara, Miyu Sasaki, Akira Nakamura, Natsumi Watanabe
Abstract
This study conducts a comprehensive analysis of the interpretability and explainability of five leading Large Language Models (LLMs): TripoSR by Stability AI, Gemma-7b by Google, Mistral 7B by Mistral AI, Llama-2-7b by Meta, and GemMoE-Beta-1 by CrystalCare AI. Through a methodical evaluation encompassing both qualitative and quantitative benchmarks, we assess these models' capacity to make their decision-making processes understandable to humans. Our findings reveal significant variability in the models' ability to provide transparent reasoning and accurate, contextually relevant explanations across different contexts. Notably, TripoSR and GemMoE-Beta-1 demonstrated superior transparency, while Gemma-7b and Llama-2-7b excelled in the accuracy of their explanations. However, challenges in maintaining consistent interpretability and explainability across varying inputs and the need for enhanced adaptability to feedback highlight areas for future improvement. This research underscores the importance of interpretability and explainability in fostering trust and reliability in LLM applications, advocating for continued advancement in these areas to achieve more transparent, accountable, and user-centric AI systems. Directions for future research include the development of standardized evaluation methodologies and interdisciplinary approaches to enhance model transparency and user understanding.