Ongoing Research Grants

Conveying Caution & Confidence: Quantification and Communication of Uncertainty in Large Language Models

Large language models (LLMs) have become immensely popular tools for laypeople and experts to find quick answers to any questions, including such that require specialized knowledge or complex reasoning. Their replies are far from always correct. Right or wrong, but often couched in a confident and persuasive tone, they may fool even experts into believing entirely false statements. Accurate and trustworthy LLM tools must know how much confidence a model has in its output and communicate this to the user in a nuanced and understandable form, respecting the principles of communication between humans. Our interdisciplinary project unites machine learning, natural language processing and linguistics. We study how LLMs acquire and verbally communicate (un)certainty, develop novel methods in Bayesian uncertainty quantification and apply them to LLMs in health applications. Using methods from linguistics, we emphasize targeted communication that gets interpreted correctly by diverse users, including laypeople and experts, and create LLMs that communicate their uncertainty according to humanlike standards.

Read more about this project in Danish or English.

This project is funded by the Independent Research Fund Denmark under the 2025 thematic call on artificial intelligence.

Project Team:

1813AI: Responsible AI for the Emergency Hotline

Every day, hundreds of citizens call the 1813 emergency hotline, but a high rate of misdirected calls puts further pressure on an already strained system. 1813AI is developing a citizen-facing, adaptive AI chat solution that, through a new self-service solution, will provide guidance and retrieve information during wait time. This creates faster access to help and contributes to a more equitable and efficient healthcare service.

This project is funded by DIREC.

Project Team:

  • Tariq Osman Andersen (PI)
  • Christian Hardmeier (co-PI)
  • Mette Bjerg Lindhøj (Region Hovedstadens Akutberedskab)
  • Lucía Gómez Zaragozá (Postdoc, ITU)
  • Hubert Dariusz Zajac (Postdoc, KU)
  • Silja Vase (Postdoc, KU)

Gradient Uncertainty: Uncertainty Quantification in Natural Language Processing

This is a PhD fellowship awarded by the Danish Data Science Academy to Nils Grünefeld. We propose a novel gradient-based framework for estimating epistemic uncertainty in machine learning models. Our approach obtains a model’s prediction, generates counterfactual outputs as hypothetical labels, and measures the resulting gradient magnitudes to assess how easily the model’s parameters would be influenced by the alternatives. The core hypotesis is that higher gradient magnitudes indicate greater epistemic uncertainty, as they reveal the model’s susceptibility to alternative possibilities. While this method is appicable to machine learning models in general, it is particularly well-suited to NLP tasks due to the natural variability in linguistic expression, which provides a rich space of meaningful counterfactuals.

Read more about this project in Danish or English.

Project Team:

  • Nils Grünefeld (PhD fellow)
  • Christian Hardmeier (Supervisor)
  • Jes Frellsen (Co-supervisor)

ALF: Enhancing Psychotherapy for Schizophrenia Spectrum Disorders using AI

The project introduces innovative use of artificial intelligence (AI) to transcribe, anonymize, and analyze Danish psychotherapy sessions for people with psychosis. By linking therapy content to treatment outcomes, it aims to uncover the core mechanisms driving therapeutic effectiveness, offering valuable insights for developing more effective interventions for patients with psychosis. The project also has the potential to expand its scope to include other diagnoses and explore additional applications of AI, ultimately aiming to enhance the accessibility, delivery and scalability of psychotherapy for a wide range of mental health disorders.

Project Team:

Past Research Grants

SafeNet: Safer Net for All

The 24-month project SafeNet: Monitoring and Reporting for Safer Online Environments (2024-2025) applied a comprehensive and intersectional approach to prevention and fight against intolerance, racism and xenophobia online. It joined 21 partners across the European Union.

In this project, ITU monitored Danish social media for hate speech and collected statistics about how platform dealt with hate speech report. I managed the Danish part of this project together with Luca Rossi after taking over from Leon Derczynski.

Project-related press releases:

  • 2024: Social media platforms fail to remove illegal hat speech (Danish, English)
  • 2023: ITU project to monitor illegal hate speech on social media platforms (Danish, English)

Neural Pronoun Models for Machine Translation

This project grant (5.2M SEK), which I got in the Swedish Research Council’s 2017 call for projects in the humanities and social sciences, supported me for 5 years before I joined ITU as an associate professor. The goal of the project was to develop computational models for the translation of pronouns in the context of neural machine translation.