ElementCheck: Complexity-Aware Factuality Evaluation for Long-Form Text Generation
Abstract
Existing approaches to long-form factuality evaluation typically rely on a decompose-retrieve-verify pipeline. However, this pipeline is susceptible to noise introduced during claim decomposition and suffers from rigid verification granularity, leading to unreliable outcomes. In response, we introduce ElementCheck, a complexity-aware framework that verifies long-form outputs through sentence-level elements. Unlike conventional methods that uniformly decompose sentences into atomic sub-claims, ElementCheck extracts entity pairs explicitly linked via verifiable connections within the original sentence, organizing them into an element graph. The graph's topology provides a structural signal for estimating sentence complexity, enabling direct verification for simple sentences and targeted element-level refinement for complex ones. To support fine-grained evaluation, we introduce FastFact-Sent, a new benchmark that maps isolated claims from FastFact-Bench back to their source sentences. Experiments on FastFact-Sent and two domain-specific benchmarks demonstrate that ElementCheck consistently improves factuality verification across five backbone models while maintaining a favorable accuracy-cost trade-off. Further analyses reveal that complexity-aware verification reduces unnecessary re-verification and maintains stability across different backbones.
1. Introduction
The rapid advancement of large language models (LLMs) has made long-form text generation increasingly prevalent across applications such as report writing, summarization, and conversational AI. Ensuring the factual accuracy of these outputs remains a critical challenge, as errors can propagate misleading information at scale. Factuality evaluation—the systematic assessment of whether generated claims are supported by reliable sources—has thus become a cornerstone of trustworthy AI systems.
Most contemporary factuality evaluation methods adopt a three-stage pipeline: claim decomposition, evidence retrieval, and verification. While effective for short-form texts, this approach struggles with long-form content where sentences are dense with interconnected facts. Decomposition often fragments meaning, introducing noise and losing contextual nuance. Moreover, fixed verification granularity fails to differentiate between simple claims that require minimal checking and complex ones that demand deeper scrutiny.
In this paper, we address these limitations by proposing ElementCheck, a novel framework that leverages sentence elements—explicitly linked entity pairs—to estimate complexity and guide verification. Our key insight is that not all sentences require equal verification effort; by analyzing structural connectivity, we can prioritize resources where they matter most.
This work makes the following contributions:
- ElementCheck: A complexity-aware verification framework that uses an element graph to estimate sentence complexity and adaptively refine verification for complex sentences.
- FastFact-Sent Benchmark: A new evaluation dataset that links claims to their source sentences, enabling fine-grained factuality assessment.
- Empirical Validation: Comprehensive experiments showing consistent improvements across five backbone models and two domain-specific benchmarks, with favorable accuracy-cost trade-offs.
2. Methodology
2.1 Element Extraction and Graph Construction
Given a source sentence, ElementCheck extracts entity pairs (e.g., subject-object or subject-predicate-object triples) that are explicitly connected through verifiable linguistic relations (such as possession, causation, or attribution). These pairs are represented as nodes in an element graph, with edges denoting dependencies or shared contexts. This structure preserves the original meaning while providing a granular view of underlying claims.
2.2 Complexity Estimation via Graph Topology
We hypothesize that graph topology correlates with sentence complexity. Simple sentences typically yield sparse graphs with few connections, whereas complex sentences produce denser structures with multiple interlinked elements. We compute several topological metrics—including node degree, clustering coefficient, and path length—to derive a complexity score for each sentence.
2.3 Adaptive Verification Strategy
Based on the complexity score, ElementCheck applies one of two verification paths:
- Direct Verification for low-complexity sentences: The entire sentence is verified against retrieved evidence as a single unit, preserving context and avoiding decomposition artifacts.
- Element-Level Verification for high-complexity sentences: The sentence is decomposed into individual elements, each independently verified, followed by a synthesis step to ensure consistency.
This adaptive approach reduces unnecessary computational overhead on simple sentences while ensuring thorough checking where complexity demands it.
3. Benchmark: FastFact-Sent
To enable fine-grained evaluation, we constructed FastFact-Sent by mapping isolated claims from the existing FastFact-Bench dataset back to their original source sentences. This mapping preserves the natural context that is often lost in claim-level benchmarks. Each sentence is annotated with gold-labeled factuality, allowing for both sentence-level and element-level evaluation.
4. Experiments
We evaluated ElementCheck against several baseline methods across FastFact-Sent and two domain-specific benchmarks (covering biomedical and legal domains). Five backbone models were tested, including both encoder-based (e.g., BERT variants) and decoder-based (e.g., T5, LLaMA) architectures.
Key Results
- ElementCheck consistently outperforms baselines in factuality verification accuracy across all benchmarks, with relative improvements ranging from 3% to 12%.
- The framework achieves a favorable accuracy-cost trade-off, reducing verification latency by up to 30% on mixed-complexity corpora due to adaptive verification.
- Complexity-aware verification significantly reduces unnecessary re-verification, particularly for simple sentences, while maintaining high recall for complex ones.
- Model stability analysis shows that ElementCheck performs robustly across all backbones, with variance in performance significantly lower than baseline methods.
5. Discussion and Conclusion
ElementCheck demonstrates that complexity-aware factuality evaluation is both feasible and effective. By moving away from uniform decomposition and leveraging structural signals, we achieve more reliable verification with better resource utilization. The introduction of FastFact-Sent provides a valuable resource for future research in fine-grained factuality assessment.
As LLMs continue to evolve in 2026, the demand for scalable and trustworthy evaluation methods grows. ElementCheck offers a principled approach that can be integrated into larger verification pipelines, potentially serving as a core component for real-time fact-checking systems. Future work will explore extending the framework to multimodal contexts and adaptive retrieval strategies based on element complexity.
Paper Information
- Authors: Xinming Wang, Haoran Du, Yi Chen, Jian Xu, Hongming Yang, Han Hu, Yulong Chen, Cheng-Lin Liu, Xu-Yao Zhang
- Submitted: 17 June 2026
- Subjects: Computation and Language (cs.CL)
- DOI: 10.48550/arXiv.2608.26118
- arXiv ID: arXiv:2608.26118
via ArXiv CL+LG