Probabilistic Chain-of-Evidence: Enhancing Factual Accuracy and Uncertainty Reasoning in Large Language Models via Prompt Engineering

Large Language Models (LLMs) frequently struggle with factual accuracy and the precise handling of uncertain information, often leading to hallucinations or misinterpretations. Existing methods like Chain-of-Thought (CoT) prompting fail to explicitly distinguish between facts and assumptions within complex contexts. To address these challenges, we introduce the Probabilistic Chain-of-Evidence (PCE) method, a novel prompt engineering strategy designed to enhance LLMs’ Factual Boundary Recognition and Uncertainty Reasoning Accuracy. PCE guides LLMs through a meta-cognitive process comprising Evidence Identification, Probabilistic Assessment, and Weighted Inference, enabling explicit quantification and integration of evidence certainty throughout reasoning. Implemented purely through sophisticated prompt design without model modifications, PCE was rigorously evaluated across diverse tasks including Factual Question Answering with Ambiguity, Medical Report Interpretation, and Legal Text Analysis. Our experiments demonstrate that PCE consistently and significantly outperforms traditional CoT prompting, achieving substantial improvements in Factual Boundary Recognition Accuracy and Uncertainty Expression Precision, while drastically reducing the Hallucination Rate. Human evaluations further corroborate these findings, indicating superior Overall Answer Quality. An ablation study confirms the crucial contribution of each PCE stage, and an analysis highlights the efficacy of a conservative “minimum” approach for robust uncertainty propagation. PCE offers a highly adaptable and practical solution for generating more reliable, transparent, and trustworthy responses from LLMs in complex, ambiguous information environments.