TriQua: Reconciling Granularity and Context in Factuality Evaluation
06:00 · August 7, 2026 · arXiv cs.AI RSS
The "decompose-then-verify" paradigm for LLM factuality evaluation faces a fundamental trade-off: atomic facts, i.e., one sentence conveying one unit of information, often omit essential context, while broader statements lack the granularity needed for precise assessment. To address this, we introduce TriQua, a framework that flexibly models facts based on their complexity. Simple claims are extracted as standard triples, while complex claims are represented as hyperrelational facts by attaching auxiliary contextual qualifiers. This adaptive structure preserves the necessary context for accurate retrieval and verification without sacrificing atomicity. Furthermore, TriQua's verification process directly annotates concrete errors within specific triples and qualifiers, providing fine-grained explainability for error detection. Alongside the framework, we propose TriQuaScore to quantify the factuality of these structured fact units. Empirical evaluations show that TriQuaScore strongly aligns with human annotated factuality scores, TriQua achieves robust decomposition quality, and outperforms existing decomposition-based frameworks in evidence-based fact verification.
Summary
TriQua addresses a core limitation in current decompose-then-verify approaches to LLM factuality evaluation. Atomic facts, typically expressed as single sentences or simple subject-relation-object triples, frequently discard modifiers that are essential for correct interpretation. At the same time, broader claims that retain those modifiers lose the granularity required for precise verification and scoring. The framework resolves this tension by adapting the representation to the complexity of each claim.
Simple assertions are extracted as conventional triples. More involved statements are modeled as base triples augmented with independent qualifiers that capture temporal, spatial, or other contextual constraints. This structure keeps each unit atomic while supplying the additional information needed for accurate evidence retrieval and checking. During verification, the system can flag errors at the level of an individual triple or a specific qualifier, rather than rejecting an entire claim when only one modifier is unsupported.
Alongside the decomposition method, the authors introduce TriQuaScore, a precision-based metric that operates directly on these structured units. Because base triples and qualifiers are scored separately, the metric supports partial credit and produces localized error annotations. Evaluations indicate that TriQuaScore correlates closely with human factuality judgments, that the decomposition step maintains high fidelity to the source text, and that the overall approach surpasses prior decomposition-based systems on evidence-grounded verification tasks.
Why it matters
This research is highly relevant for Dutch AI researchers and practitioners focused on trustworthy AI and LLM deployment. Improving factuality evaluation directly supports the Netherlands and EU strategic emphasis on transparent, reliable, and ethical AI systems.
