EVIDENZ: Evidence-based text generation using LLMs

Name of participant: Tobias Schreieder

Project’s name: Evidenzbasierte Textgenerierung mit LLMs (EVIDENZ)

Project description:

The “Evidence-Based Text Generation with LLMs” (EVIDENZ) project aims to enhance the trustworthiness of large language models (LLMs) in scientific application scenarios. The focus is on developing methods for evidence-based text generation, in which generated content is systematically linked to traceable and verifiable references. This is intended to enable users to independently validate statements made by LLMs using scientific sources.

EVIDENZ addresses key challenges faced by current large language models (LLMs), which, whilst often providing linguistically convincing answers, frequently fail to provide any citations or provide only inadequately relevant ones. To this end, the project is investigating both non-parametric approaches that integrate external academic sources at runtime, and parametric methods that enable the traced back of generated content to training data.

The partnership with Springer Nature significantly enhances the practical relevance of the project. As one of the world’s leading scientific publishers, Springer Nature brings extensive expertise as well as access to high-quality scientific publications and metadata. These resources make it possible to address realistic application scenarios and evaluate the developed methods under real-world conditions.

A key component of the project is the development of a novel scientific benchmark for evidence-based text generation using large language models (LLMs). This benchmark is designed to systematically evaluate the quality of generated content and citations, and to enable a comparison of different approaches. Building on this, a prototype for Retrieval-Augmented Generation (RAG) is being developed, which specifically selects relevant sources and integrates them into the generation process, taking particular account of the citation function (the purpose of a citation). In addition, parametric methods for attribution are being analysed and further developed with regard to their scalability and applicability in an academic context.

Software Campus Partner: Technische Universität Dresden und Holtzbrinck

Implementation period: 01.01.2026 – 31.12.2027