Large Language Models for Research Data Management?! 2025 (LLMs4RDM 2025)

Abstract

Research data management (RDM) has become an important discipline that enables researchers to efectively organise, preserve and share their research results. RDM is a new development that aims to prepare researchers for the future by building on the principles of open science. It utilises innovative approaches such as generative artificial intelligence (genAI), which is powered by large language models (LLMs), to complement traditional research methods. As data-driven research becomes increasingly complex, researchers often have to spend a lot of time learning how to manage, analyse and interpret large amounts of information. Traditional data literacy training can be time-consuming and doesn't always keep pace with evolving technologies and methods of analysis. Foundation models based on generative AI ofer the potential to streamline this learning process. By automating data pre-processing, pattern recognition and even hypothesis generation, these models can lower the technical barriers to entry, allowing researchers to focus more on insights and discovery rather than spending excessive amounts of time mastering data skills. The objective of this workshop is an exchange of perspectives regarding the implementation of novel RDM approaches using LLMs or not, both past and prospective, in research and practice.

1.1. Workshop Organisation

The LLMs4RDM 2025 Workshop was held as part the INFORMATIK Festival 2025 (55th Annual Conference of the German Informatics Society), September 18, 2025, Potsdam, Germany. 1.1.1. Organisers • Magnus Bender, Aarhus University, Denmark • Sylvia Melzer, University of Hamburg, Germany • Ralf Möller, University of Hamburg, Germany • Stefan Thiemann, University of Hamburg, Germany 1.2. Programme Committee of LLMs4RDM 2025 • Thomas Asselborn, University of Hamburg, Germany • Magnus Bender, Aarhus University, Denmark • Mahdi Jampour, University of Hamburg, Germany • Sylvia Melzer, University of Hamburg, Germany • Stefan Thiemann, University of Hamburg, Germany

1.3. Overview of papers

One keynote and five papers were presented at the workshop.

The keynote focuses on the crucial next step in Research Data Management (RDM), advocating for a transition from simple data immersion to structured scientific argumentation. The speaker presents historical examples, such as the 3D reconstruction of the theatre of Miletus, to illustrate how researchers formulate hypotheses about past human decisions, pointing out that these visualisations can be based either on empirical data or on imaginary 3D data generated via language models. The central argument is that RDM systems must evolve to support the formal representation of these scientific arguments so that they are machine-processable and the data used can be verifiably represented. The keynote calls for new RDM systems that not only host diverse data, but also enable researchers to use these resources directly for the development and validation of formal scientific hypotheses.

The first paper Large Language Models in Labor Market Research Data Management: Potentials and Limitations presents the application of LLMs within research data management (RDM), focusing specifically on tasks related to occupational data and labor market text interpretation. Through empirical studies, the researchers determined that LLMs struggle with the automated classification of job titles, often producing results that were less reliable and reproducible than those generated by traditional machine learning classifiers. The LLM tests using hermeneutical methods produced fundamentally inconsistent and unstable interpretations. The authors argue that LLMs are inadequate for tasks demanding methodological rigor or scientifically defensible classification due to their lack of consistency and interpretative depth. Therefore, LLMs should be used only as assistive tools for preliminary support functions.

The second paper Challenges in Automatic Speech Recognition in the Research on Multilingualism examines the significant challenges faced when applying Automatic Speech Recognition (ASR) technology, specifically the Whisper model, to complex spoken data collected for multilingualism research.

The authors note that while commercial applications require clean, monolingual transcripts, linguistic

studies require highly accurate recordings that capture every acoustic detail, including speech disorders and complex language switching between languages. Using Polish-German bilingual recordings from the LangGener corpus, the study identifies key shortcomings in ASR output, such as hallucinations and a problematic tendency towards code unification, which mis-transcribes or mis-translates embedded language elements.

The third paper Improving Accessibility and Reproducibility by Guiding Large Language Models presents proposes a combined method the general-purpose large language models (LLMs) and specialized research data stored in Research Data Repositories (RDRs) by leveraging the expert knowledge of the data creators.

The core innovation is the interpretation prompt, a field added during the data upload process that

allows the expert data creator to provide specific instructions. When a user queries the RDR’s LLM chatbot, this expert-generated prompt is prepended to the user’s query, efectively guiding the LLM toward project-specific understanding. The authors demonstrate that these prompts result in more accurate, tailored responses by focusing the LLM’s output and improving data accessibility and utility.

Furthermore, the interpretation prompt facilitates automated reproducibility of research experiments by instructing the LLM to execute relevant algorithms or code associated with the data entry.

The forth paper Talk to your Database: An open-source in-context Learning Approach to interact with

Relational Databases through LLMs presents an open-source large language model (LLM) framework

designed to solve the Text-to-SQL problem through in-context learning. Researchers compared the performance of this method against a simpler default prompting technique using a PostgreSQL database.

The results decisively show that in-context learning improved accuracy, boosting successful query execution rates from approximately 35% to over 85%.

The fifth paper Verbalisation Process of a RAG-Based Chatbot to Support Tabular Data Evaluation for Humanities Researchers presents the verbalization process of a RAG-based chatbot (ChatHA) engineered to support tabular data evaluation for humanities researchers. The core motivation for this research is enabling scholars to conduct free-form and semantic searches on structured data, moving beyond the limitations of simple string matching. To address the need for verbalizing database entries into natural language, the authors propose a hybrid verbalization method that minimizes the computational cost and risk of hallucination associated with LLMs.

2. Presentations

2.1. Presentations

Magnus Bender

Aarhus University, Denmark Welcome

Ralf Möller

University of Hamburg, Germany Keynote

Abstracts and presentations are available at: https://doi.org/10.25592/uhhfdm.17955

Jens Dörpinghaus1,2,3, Michael Tiemann1,2 1University of Koblenz, Germany; 2Federal Institute for Vocational Education and Training (BIBB), Germany; 3Linnaeus University, Sweden Large Language Models in Labor Market Research Data Management: Potentials and Limitations Edyta Jurkiewicz-Rohrbacher1,2, Thomas Asselborn2 1University of Regensburg, Germany; 2University of Hamburg, Germany Challenges in Automatic Speech Recognition in the Research on Multilingualism

Florian Marwitz, Marcel Gehrke

University of Hamburg, Germany Improving Accessibility and Reproducibility by Guiding Large Language Models

Maximilian Plazotta, Meike Klettke

University of Regensburg, Germany Talk to your database: An open-source in-context learning approach to interact with relational databases through LLMs Thomas Asselborn1, Magnus Bender2, Florian Marwitz1, Ralf Möller1, Sylvia Melzer1 1University of Hamburg, Germany; 2Aarhus University, Denmark Verbalisation Process of a RAG-Based Chatbot to Support Tabular Data Evaluation for Humanities Researchers

Magnus Bender

Aarhus University, Denmark Farewell 2.1.1. Acknowledgments

The organisers of the LLMs4RDM 2025 workshop would like to thank the organisers of the Informatik

Festival conference in Potsdam for their excellent support. We also would like to thank the members of the programme committee for their help in carefully evaluating and selecting the submitted papers and all participants of the workshop for their contributions. We wish that new inspirations and collaborations between the contributing disciplines will emerge from this workshop.

Funding Information

This contribution was partially funded by the Deutsche Forschungsgemeinschaft (DFG, German Research

Foundation) under Germany´s Excellence Strategy – EXC 2176 ‘Understanding Written Artefacts:

Material, Interaction and Transmission in Manuscript Cultures’, project no. 390893796. The research

was mainly conducted within the scope of the Centre for the Study of Manuscript Cultures (CSMC) at

University of Hamburg.

This contribution was partially funded by the Danish National Research Foundation (DNRF193) through TEXT: Centre for Contemporary Cultures of Text at Aarhus University.

Declaration on Generative AI

During the preparation of this work, the authors used DeepL in order to: Grammar and spelling check.

After using these tool(s)/service(s), the authors reviewed and edited the content as needed and take(s)

full responsibility for the publication’s content.

References