EVIDENZ

Evidence-based Text Generation with LLMs

Large language models are increasingly part of how researchers search, read, and write.
At the same time, scientific information overload keeps growing, and it becomes harder to judge which claims are actually supported by evidence.

EVIDENZ develops novel AI methods for evidence-based text generation with LLMs in the scientific domain.
The goal is straightforward: answers should come with suitable references, and readers should be able to verify the key claims.

What we do

We pursue three main research directions:

  1. Scientific QA benchmark
    We build a dataset with question–answer–evidence triples that captures not only what to cite, but also why a reference is used (e.g., background, comparison, support).

  2. Retrieval-augmented generation prototype
    We develop a retrieval pipeline that deliberately selects evidence that matches the intended citation function, and we combine it with generation that produces answers with supporting references.

  3. Parametric attribution analysis
    We study approaches that link generated content back to training sources and analyze their potential and limitations in realistic, scalable settings.

Why it matters

Science is already suffering from information overload: too many papers, too little time.
AI assistants can help researchers navigate this, but only if they make it clear what is evidence and what is generation.

Today’s LLMs can produce fluent answers even when the underlying support is weak, irrelevant, or missing.
In scientific workflows this is a serious problem: users need to trust that key statements can be traced back to concrete sources.

EVIDENZ targets this gap by treating traceability and provenance as core requirements.
We push beyond standard retrieval-augmented generation by developing new methods for evidence selection, citation-function modeling, and attribution.

Get involved

If you are interested in working on this project:

  • Students: thesis / HiWi opportunities are available — see /student_thesis/ or contact Tobias Schreieder.
  • Media: for expert interviews on AI reliability, citations, chatbots, and scientific QA, please see /media/.

Lead: Tobias Schreieder
Academic supervisor: Michael Färber
Industry partner: Holtzbrinck / Springer Nature
Start: 01/2026 · Duration: 24 months

References