Quantifying Factual Divergence in Generative Models: SHAP-LIME Based Hallucination Score for LLMs

Key Points

• Hallucination scores quantify discrepancies in outputs of large language models, enhancing evaluation methodologies.
• The defined score allows a clearer insight into generative models' performance, especially in misinformation contexts.
• Assessments using SHAP and LIME techniques elucidate model behavior, forecasting improvements for future LLM applications.
• Findings highlight the need for standardized metrics in evaluating generative models to ensure reliability.