New Framework Improves LLM Factuality Evaluation with Context.

Evaluating the factuality of Large Language Models (LLMs) often relies on a "decompose-then-verify" approach, where statements are broken down into atomic facts for assessment. However, this method faces a fundamental dilemma: atomic facts, while precise, frequently lack the necessary context for accurate verification, while broader statements lose the granularity needed for precise error identification. This trade-off makes it difficult to achieve both accuracy and explainability in factuality evaluations. To overcome this, researchers have introduced TriQua, a flexible framework designed to model facts based on their inherent complexity. For simple claims, TriQua extracts standard triples (subject-predicate-object). For more complex claims, it represents them as "hyperrelational facts" by attaching auxiliary contextual qualifiers. This adaptive structure ensures that crucial context is preserved for reliable retrieval and verification without sacrificing the atomicity required for precise assessment. Beyond its flexible fact representation, TriQua's verification process directly annotates specific errors within individual triples and qualifiers, offering highly granular and explainable error detection. The framework also includes TriQuaScore, a metric designed to quantify the factuality of these structured fact units. Empirical evaluations demonstrate that TriQuaScore exhibits strong alignment with human-annotated factuality scores, and the framework achieves robust decomposition quality, outperforming existing decomposition-based methods in evidence-based fact verification.