Can AI understand culture, biology and irony? Four research projects receive millions in funding

Principal investigator: Tenure-track Assistant Professor Daniel Herscovich  

Co-principal investigator: Associate Professor Jens Bjerring-Hansen (Department of Nordic Studies and Linguistics) 

Grant amount: DKK 7,199,998 

Abstract: 
This project develops new methods for evaluating and improving cultural reasoning in language models using literary texts. While current AI systems can retrieve facts, they struggle with interpretive tasks such as irony, perspective, and cultural references—especially in underrepresented languages like Danish. CuRe addresses this gap by creating soft, human-annotated benchmarks grounded in literary analysis, and by testing models such as retrieval-augmented generation (RAG) systems on these tasks. The project integrates insights from NLP, literary studies, and humanistic evaluation theory, advancing culturally responsible AI. It involves partners at UC Berkeley, McGill, and the University of Copenhagen, and contributes to international efforts in fair and explainable language technologies. 

Principal investigator: Professor Isabelle Augenstein 

Co-principal investigator: Professor Irina Shklovski 

Grant amount: DKK 7,196,132 kr. 

Abstract: 
Despite their high performance, large language models (LLMs) exhibit many factual errors. Additional documents are commonly retrieved to increase an LLM's knowledge, but an LLM may ignore these documents when generating text. There is limited work to illuminate these inner workings to users and to promote interventions and a more in-depth understanding of these issues. This project will develop explainable AI methods to increase fact-based learning education and applicability to information retrieval by enabling physicians to justify generated output and enabling interventions based on applied evidence. It will generate new insights into strategic evidence gathering and knowledge management, and what users consider to be key factors for critical evaluation of LLM output. 

Principal investigator: Professor Wouter Boomsma 

Co-principal investigator: Dr. Antonio Vergari (University of Edinburgh) 

Grant amount: DKK 6,572,380  

Abstract: 
Large AI models are transforming biology. For instance for proteins, language and structure-to-sequence models have fundamentally altered the analysis of evolutionary patterns and design of new variants. However, critical prediction tasks in biology remain out-of-reach for AI because we lack principled ways to enforce constraints, such as codon structure in gene annotation or structural constraints from knowledge graphs. Neuro-symbolic (NeSy) machine learning offers a promising solution by combining neural networks’ expressiveness with symbolic AI’s reasoning. Yet symbolic methods often lack scalability and are hard to integrate with large neural models. We propose a NeSy framework enabling tractable inference, including exact conditioning, to produce flexible, biologically grounded predictions. By extending tractable NeSy models such as circuits and integrating them with biomolecular foundation models, we aim to create robust, interpretable and biologically faithful systems. 

Principal investigator: Professor Anders Krogh 

Participant: Research Assistant Adrián Sousa-Poza will be employed as a PhD student 

Grant amount: DKK 3,024,592 

Abstract: 
This project aims to develop deep generative models that understand gene regulatory networks (GRNs) and predict cellular responses to perturbations of gene expression. By training on single-cell data, these models will learn regulatory mechanisms and be able to predict unseen cell perturbations. The research builds upon the Deep Generative Decoder (DGD) framework developed by the Krogh group combined with novel deep learning methods like transformers and diffusion models. The project addresses key challenges, including data limitations, interpretability gaps, and standardisation needs. Through collaboration, we will have access to unique multi-omic and time series datasets crucial for model validation. This work advances in silico perturbation modelling by integrating multi-omics data and leveraging GRN information, potentially  reducing reliance on costly experiments and contributing to more personalised treatment approaches for complex diseases.