Reliable AI analyses for business decisions: Joint Graduate Research Training Group of Goethe University, TU Darmstadt, and Leibniz Institute SAFE

Bild: Adobe Stock / itchaznong

Bild: Adobe Stock / itchaznong

Today, Artificial Intelligence (AI) can evaluate not only tables but also texts, images, audio, and video recordings. This holds great potential for companies, financial markets, and regulatory authorities—provided the results are reliable, traceable, and economically interpretable. A new Research Training Group (Graduierenkolleg) proposed by Goethe University and TU Darmstadt brings together computer science and economics to this end: it develops methods for such multimodal analyses and tests them on concrete issues, ranging from patent drawings and annual reports to greenwashing.

Whether annual reports, patent specifications, sustainability reports, or recordings of analyst conferences: a large portion of economically relevant information is not available in tables, but exists as unstructured text, graphics, images, audio, or video. While modern AI models can process such data, they face limitations in practice: Large Language Models (LLMs) are difficult to apply to very large datasets, they occasionally “hallucinate,” and it often remains unclear how a result was reached and how robust it is. In the context of financial markets, sustainability reporting, or regulation, this is unacceptable.

This is where the Graduate Research Training Group “Advanced Content Analytics for Corporate Decision-Making and Regulation” comes in. The program was approved today by the German Research Foundation (DFG). It was jointly applied for by Professor Oliver Hinz (Economics) of Goethe University Frankfurt and Professor Iryna Gurevych (Computer Science) of TU Darmstadt. The group integrates three research areas: In the first, computer scientists develop the technological foundations—namely scalable, reliable, interpretable, and transparent analysis procedures. In the second research area, these procedures are applied to economic and regulatory questions. The third area synthesizes the findings across projects and derives general conclusions.

Measuring economic phenomena more accurately

“We are deliberately working in both directions: economic and regulatory questions present computer science with new methodological challenges, and conversely, the new methods allow us to measure economic phenomena differently and more accurately than ever before,” explains Professor Oliver Hinz, spokesperson for the Graduate Research Training Group and Professor of Information Systems and Information Management at Goethe University Frankfurt. “For us, an answer from an AI model is not yet 'information.' It only becomes so once it has been verified, its uncertainty quantified, and it can be interpreted economically.”

One of the four computer science projects demonstrates how this will work in practice: the goal is to create an “execution engine” for multimodal data—a system that executes queries across large volumes of texts and images. For every analysis step, various versions of AI models are available, which operate with different speeds and levels of accuracy. An optimizer will automatically select the fastest execution plan that maintains a minimum accuracy specified by the user. The uncertainty of individual steps is statistically quantified and carried through the entire analysis chain. This makes it possible to demonstrate how robust a result is. Other computer science projects will investigate, for example, how feedback from experts such as financial analysts can be systematically integrated into AI-supported analyses, or how AI systems can adapt their approach to the specific type of document and verify every analysis step.

Application projects provide real-world data

Six research projects from the field of economics provide the research questions and real-world data, serving later as practical tests for the new procedures. They examine, for example, how technical drawings influence the comprehensibility, granting, and litigiousness of patents; what role images and graphics in annual reports play for shareholders; and how culture and trust in the financial sector can be measured using audio and video recordings from companies. Further projects ask, for instance, what impact sustainability reporting by non-listed companies in the EU has and how greenwashing can be uncovered by comparing sustainability claims from companies and funds with environmental data.

Starting in the third year, five “integration clusters” will address questions that span all projects—such as how AI models handle the fact that language and data change over time, what added value additional data types provide, or when specialized models are superior to general language models. Doctoral students will be supervised in tandems consisting of experts from both disciplines. A joint core curriculum will provide the fundamentals; a lecture series, international workshops, and research stays abroad are also planned.

More efficient and trustworthy collaboration

Professor Gurevych’s Ubiquitous Knowledge Processing (UKP) Lab brings expertise to the Research Training Group in human-AI collaboration and in aligning large language models (LLMs) with expert feedback and preferences. “Within the Research Training Group, we are developing a framework that turns diverse expert feedback – such as corrections, rankings, critiques and demonstrations – into business analysis artefacts, such as analyst ratings and argumentative reports,” explained Gurevych, deputy spokesperson for the Research Training Group. “Analysts and models co-construct these from heterogeneous sources such as fillings, news, market time series, audio recordings of earnings calls or charts. This allows faster, more trustworthy collaboration, while humans keep control and the final sign-off.”

The framework builds on work carried out by the UKP on interactive annotation, multi-document reasoning, argument mining and reward modelling as part of the ERC Advanced Grant and the LOEWE projects. At TU Darmstadt, in addition to the UKP Lab, Professor Carsten Binnig’s Systems Group and Professor Marcus Rohrbach’s Multimodal Reliable AI research group (both also part of the Department of Computer Science) are involved.

15 doctoral positions planned

In addition to the applying universities—Goethe University Frankfurt as the lead institution and the Technical University of Darmstadt—the Leibniz Institute for Financial Market Research SAFE is also involved in the group. The first funding period begins on April 1, 2027, and will run for five years, with a total of 15 doctoral positions planned—ten at the start and five more starting in 2029. Additionally, three associated postdocs will participate, though they will not be funded by the group’s resources. A four-year extension of the program is possible.

Ten professors are involved in the group. “The doctoral positions will likely be advertised in December 2026 or January 2027. We are pleased to bring the ideas of the RMU-EXCITE Excellence Initiative to life in two profile areas, 'Data' and 'Society,' through this joint Graduate Research Training Group,” says Oliver Hinz.

“Within the Research Training Group, the contribution of each RMU partner strengthens the others,” emphasizes Gurevych, Professor of Ubiquitous Knowledge Processing. “The expertise, analytical data and domain understanding at Goethe University and SAFE enable TU Darmstadt to develop and test cutting-edge AI methods on such real problems. In return, our business partners gain feedback-driven AI tools for challenging domain-specific questions in their own empirical research. Doctoral researchers on both sides gain interdisciplinary co-supervision, shared RMU courses and access to the regional AI ecosystem.”

Text: GU / TU

Rhine-Main Universities