Overview of PRESTA at IberLEF 2025: Question Answering Over Tabular Data In Spanish

Resumen

We present the findings and results of the PRESTA track at IberLEF 2025, focused on question answering over tabular data in Spanish. The task challenges participants to build systems capable of interpreting natural language questions and retrieving accurate answers from semi-structured tabular sources in Spanish. In this paper, we describe the task design, dataset construction, evaluation methodology, and participant systems. We analyze a range of submitted approaches and discuss key trends observed across systems. Our results show that methods leveraging large language models (LLMs) clearly outperformed traditional pipelines, with larger multilingual models exhibiting very high accuracy. It is of note that the performance of small open-source models is up to par with the bigger proprietary ones when paired with good system designs. These findings confirm that the strong performance of LLMs in English carries over to Spanish in the context of tabular question answering, though some linguistic and domain-specific challenges remain.