@article{cag2026sachdeva,
title = {Systematic validation of LLM-generated structured data - A design space and remaining challenges},
journal = {Computers & Graphics},
volume = {135},
pages = {104545},
year = {2026},
issn = {0097-8493},
pdf = {https://papers.ssrn.com/sol3/Delivery.cfm?abstractid=5626640},
doi = {https://doi.org/10.1016/j.cag.2026.104545},
url = {https://www.sciencedirect.com/science/article/pii/S0097849326000166},
author = {Madhav Sachdeva and Christopher Narayanan and Marvin Wiedenkeller and Jana Sedlakova and Jü\"{u}rgen Bernard},
keywords = {Large language models, Validation, Design space},
abstract = {Large language models (LLMs) are increasingly being used in academia and practice to generate structured data, supporting crucial data enrichment tasks such as imputing missing values, labeling data items, and generating synthetic datasets. However, these benefits rely on the validation of LLM-generated data to address known issues of LLMs, including hallucinations, inconsistencies, logical contradictions, and biases. Despite its importance and the significant growth of validation approaches in both diversity and count, the space opened up by these validation approaches remains unstructured. Based on a systematic literature review, we present a design space for approaches to the validation of LLM-generated structured data. The design space structures these approaches along two primary dimensions: Data Source and Granularity, and extends them with three complementary dimensions: Visualization techniques, Interaction techniques, and Workflow phases. Together, these dimensions form the descriptive, evaluative, and generative power of the design space. We apply the design space to demonstrate its utility through the analysis of three representative LLM-based validation approaches for structured data. Moreover, we reflect on the development process of Val-LLM, an interactive visual tool for multi-granularity validation, leveraging the design space as guideline in a novel approach. The results show that the design space enables researchers and practitioners to systematically characterize validation methods and guide the design of interactive systems for validation. We conclude by discussing limitations, remaining challenges, opportunities to extend the design space and to advance future validation research and practice.},
teaserpage = {1},
topics = {Explainable and Trustworthy AI, Visual-Interactive Labeling (VIAL), Theoretical Foundations of Visualization},
code = {J041}
}