Researchers propose ElementCheck, a complexity-aware framework for long-form text factuality evaluation

A new framework called ElementCheck has been proposed to address challenges in long-form text factuality evaluation. Unlike existing pipelines that rely on decompose-retrieve-verify methods, ElementCheck extracts entity pairs explicitly linked through verifiable connections in the original sentence as elements. These elements are organized into an element graph, where graph topology provides a structural signal for estimating sentence complexity.

The framework adapts verification granularity based on sentence complexity: simple sentences undergo direct verification, while complex sentences receive targeted element-level refinement and verification. This approach aims to reduce noise from claim decomposition and fixed verification granularity, which have been identified as sources of unreliable results in prior methods.

To support fine-grained evaluation, the authors construct a new benchmark named FastFact-Sent. This benchmark maps isolated claims from FastFact-Bench back to their source sentences, enabling more precise assessment of factuality at the sentence level.

In experiments conducted on FastFact-Sent and two domain-specific benchmarks, ElementCheck consistently improves factuality verification across five backbone models. The framework also maintains a favorable accuracy-cost trade-off, and further analyses indicate that complexity-aware verification reduces unnecessary re-verification while maintaining stability across different backbones.