Confident but Conflicted: Internal Uncertainty and Cognitive Dissonance Resolution in LLMs

The paper introduces the concept of Trust Elasticity (TE), a measure of how large language models (LLMs) resolve cognitive dissonance when faced with conflicting evidence. The study evaluates TE across four LLMs, revealing that it varies significantly with source authority and evidence quality, while also correlating with internal uncertainty indicators such as Confidence Miscalibration in Qwen and Internal Uncertainty Change in Llama. This research highlights the importance of understanding internal model uncertainty for improving LLM responses to conflicting information, suggesting potential avenues for enhancing model reliability in practical applications.