Best Paper Award for IVDA at EuroVA 2025
The paper “A Design Space for the Critical Validation of LLM-Generated Tabular Data” by Madhav Sachdeva, Christopher Narayanan, Marvin Wiedenkeller, Jana Sedlakova, and Jürgen Bernard has been honored by the Best Paper Award at EuroVA 2025 in Luxembourg.
EuroVA is the largest workshop at the EuroVis Conference, Europe’s largest conference for interactive visual data visualization, hosted by the Eurographics Association.
Congratulations to the IVDA team!
Abstract
Large language models (LLMs) – the technology behind tools like ChatGPT – can now generate realistic-looking tables of data. This opens up exciting new possibilities for researchers, businesses, and society. These tables can help fill in missing information, add useful labels, or include extra details based on context.
However, before we can fully trust and use this AI-generated data, we need to critically validate its correctness. Right now, many new methods are being developed to do this kind of validation, but the field is still a bit disorganized.
To bring more structure, we introduce a Design Space that helps people think about how to validate LLM-generated tables. It has two main parts:
- Level of Detail: Are we checking individual data points, groups of values, or how different columns relate to each other?
- Source of Information: Are we comparing the AI’s output to real-world data, explanations, or a mix of both?
We explain how different types of analysis fit into this Design Space, review 19 current methods, and take a closer look at two of them to show how powerful this approach can be.