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

Researchers introduce CREDENCE, a new framework for decomposing complex claims into verifiable atomic statements, addressing limitations in existing fact-checking pipelines. The framework replaces token-overlap metrics with semantic similarity scoring and provides formal convergence analysis for repair loops, improving fact-checking accuracy by 15-32 percentage points across multiple domains.

CREDENCE tackles a fundamental challenge in automated fact-checking: the reliable decomposition of compound claims into simpler, verifiable units. Traditional approaches using Jaccard similarity metrics systematically fail when claims are paraphrased or semantically equivalent, leading to underestimated decomposition quality and cascading errors in downstream fact-checking systems. This research addresses a critical infrastructure gap in computational fact-verification that affects both academic research and real-world deployment of fact-checking tools.

The framework's key innovation replaces token-overlap metrics with BGE-large cosine similarity, a semantic fidelity measure that better captures meaning-preserving transformations. Beyond methodology, the authors provide formal mathematical proofs establishing convergence properties of their repair pipeline—a typically overlooked aspect of claim decomposition systems. Rule-based repair is proven monotone and finitely terminating, while LLM-based self-repair requires safeguards against infinite loops, reflecting practical tensions between automation and reliability.

The research demonstrates meaningful performance gains across three distinct evaluation benchmarks spanning social media, encyclopedic, and news domains. Semantic-F1 outperforms traditional metrics by substantial margins, while atomicity violation rates drop 47-100% with repair mechanisms. These results have direct implications for developers building fact-checking systems, as they can adopt more robust evaluation metrics and termination guarantees. The multi-model benchmarking across 3.8B-12B parameter models and API-based systems provides actionable guidance for practitioners choosing decomposition approaches.

Looking forward, the formal convergence framework establishes a foundation for more reliable claim decomposition at scale. Future work should explore how these semantic metrics and termination guarantees transfer to multilingual fact-checking and handle adversarially constructed claims designed to resist decomposition.

• →Semantic-F1 outperforms Jaccard metrics by 15-32 percentage points in claim decomposition tasks
• →CREDENCE provides formal proof that rule-based repair is monotone and finitely terminating
• →LLM-based self-repair requires early-exit guards to prevent non-monotone behavior and infinite loops
• →Framework achieves 0.94-1.00 exact pair recall on social media and encyclopedic benchmarks
• →Multi-domain evaluation across social media, Wikipedia, and news sources demonstrates cross-domain generalization