CREDENCE: Claim Reduction for Decomposition & Enhanced Credibility -- Semantic Metrics and Convergence Analysis

The article introduces CREDENCE, a claim decomposition and evaluation framework that enhances automated fact-checking by addressing limitations of previous methods. It features a new Semantic-F1 metric based on BGE-large cosine similarity, demonstrating a significant performance improvement over Jaccard metrics, with benchmark results showing increases of 15-32 percentage points in accuracy. Additionally, the framework provides formal convergence theorems for the repair pipeline and includes multi-model evaluations across various domains, making it a valuable tool for practitioners aiming to improve the reliability of fact-checking systems using LLMs.