TriQua adds context qualifiers to make factuality checks less brittle

TriQua proposes a way to make LLM factuality evaluation less brittle by preserving context without giving up fine-grained checking.

Many factuality systems decompose generated text into atomic claims and then verify them. The paper argues that this creates a tradeoff: very small claims can lose the context needed for accurate verification, while broader statements are too coarse to assess precisely.

TriQua handles simple claims as standard triples and complex claims as hyperrelational facts with contextual qualifiers. Its verification process can then annotate concrete errors inside specific triples or qualifiers. The authors also propose TriQuaScore, which they say aligns strongly with human factuality judgments and outperforms existing decomposition-based frameworks in their evaluations.

The practical point is that factuality tools need to know what exactly is being checked. A claim without context can be misleading, but a claim with too much bundled information is hard to score. TriQua tries to balance both needs.