CuRe: Cultural Reasoning for Responsible Language Model Development
CuRe develops new evaluation methods to test and improve how language models interpret culture by using Danish literature as a rigorous ground for cultural reasoning.
CuRe: Cultural Reasoning for Responsible Language Model Development
Independent Research Fund Denmark (DFF), 2026–2030
Project Summary
CuRe investigates how language models interpret culture — not through facts or stereotypes, but through the rich, ambiguous, and historically layered world of literature. By combining NLP, literary studies, and expert-driven evaluation design, the project develops new methods for assessing and improving cultural reasoning in AI systems.
Why Culture? Why Literature?
Culture shapes meaning, and meaning is where AI struggles most.
Literature embeds cultural knowledge through:
- Referencing — allusions, echoes of earlier works
- Reuse — motifs, clichés, idioms, recurring patterns
- Perspective and interpretation — multiple legitimate readings
This makes literature the ideal testing ground for evaluating how AI understands culture — and for building models that respect cultural nuance rather than reducing it to stereotypes.
Objectives
CuRe addresses three core research questions:
- How can we build robust benchmarks for cultural reasoning where multiple interpretations are valid?
- How do we model and evaluate interpretive depth in AI, beyond surface pattern recognition?
- How can we improve cultural reasoning in AI responsibly, without reinforcing essentialism?
Work Packages
WP1 — Empirical Foundations
Construction of high-quality corpora and interpretive benchmarks based on Danish literature, including MeMo, Mini-WorldLit, and canonical texts. Data includes passages, interpretive annotations, student essays, and expert commentary, designed following the ECBD framework.
WP2 — Methodological Reflection
Evaluation of retrieval-augmented generation (RAG), long-context models, and soft-label annotation strategies. Analysis of interpretive ambiguity, multiple valid readings, and the relationship between close and distant reading.
WP3 — Model Adaptation & Training
Adaptation of Danish Foundation Models (DFM), experiments with pretraining mixtures, fine-tuning with expert feedback, and human-in-the-loop reinforcement learning.
Team
Principal Investigators
-
Daniel Hershcovich (PI) Tenure-Track Assistant Professor, Department of Computer Science, University of Copenhagen
-
Jens Bjerring-Hansen (Co-PI) Associate Professor, Department of Nordic Studies and Linguistics, University of Copenhagen
-
Desmond Elliott (Advisor) Associate Professor, Department of Computer Science, University of Copenhagen
Researchers
- Alexander Conroy (Postdoc)
- PhD Student (2027–2030)
International Collaborators
- David Bamman (UC Berkeley)
- Timothy Tangherlini (UC Berkeley)
- Maria Antoniak (University of Colorado)
- Jackie Cheung (McGill University)
Advisory Board
- Serge Belongie
- Anders Søgaard
- Carsten Levisen
- Bolette Sandford Pedersen
Methods and Approach
- Expert-driven annotation of interpretive phenomena
- Soft labels capturing multiple valid interpretations
- Retrieval-augmented reasoning over literary corpora
- Human-in-the-loop refinement of model behaviour
- Comparisons of open-source and proprietary LLMs
- Transparent benchmark design via ECBD principles
Expected Outcomes
- A publicly released benchmark suite for cultural and literary reasoning
- New corpora and annotations grounded in Danish literary heritage
- Methodological insights into responsible cultural evaluation
- Improved AI reasoning about cultural and historical context
- Cross-disciplinary impact across NLP and the humanities
Publications
A rolling list of project publications (2026–2030) will be maintained here.
- Benchmark releases
- Methodological papers
- Joint CS–literature analyses
- Workshops and conference outputs
Contact
For inquiries or collaboration:
dh@di.ku.dk