Navigating Unreliable Parametric and Contextual Knowledge: Explicit Knowledge Conflict Resolution for LLM Inference

A new framework called MACR has been proposed for resolving conflicts between parametric and contextual knowledge in large language models (LLMs). This framework utilizes an adaptive knowledge assessment approach based on a modified semantic entropy measure to evaluate the model's confidence and employs an inductive multi-agent reasoning system to analyze and resolve inconsistencies among various knowledge sources. MACR has shown significant improvements over existing state-of-the-art methods in empirical benchmarks, enhancing the reliability of LLMs in scenarios where both internal and external information may be erroneous.