Evaluating Consistencies in LLM responses through a Semantic Clustering of Question Answering

LLMs produce inconsistent answers to identical questions, which undermines user trust—this study quantifies this problem by generating multiple responses to the same questions and measuring how often the semantic meaning differs, even when the wording varies.

Two proven techniques can improve consistency: Retrieval-Augmented Generation (RAG), which provides external knowledge context, and Zero-shot Chain-of-Thought (CoT), which prompts the model to reason through problems step-by-step, both demonstrably enhance response reliability across different types of questions.

Semantic similarity matters more than exact word matching—the research recognizes that different phrasings can convey identical meanings, so evaluating consistency requires looking beyond surface-level text differences to understand whether responses actually say the same thing.

The problem is widespread across domains—testing on 37 different question categories from the TruthfulQA dataset shows that inconsistency isn't limited to one type of question or subject area, suggesting this is a fundamental challenge in how LLMs currently operate.

Randomness in how LLMs generate text is a root cause—the inconsistency stems from the probabilistic nature of token sampling (the random selection process LLMs use to generate each word), indicating that improving consistency may require changes to how these models fundamentally work.