Independent Research Fund Denmark · 2026–2030

Cultural Reasoning for Responsible Language Model Development

How good are LLMs at cultural reasoning? Can we make them better? What does “better” even mean?

CuRe brings together natural language processing and literary studies to develop better ways of evaluating and training language models for culturally grounded reasoning.

PhD position Publications

We are hiring

PhD fellowship in Cultural Reasoning with Language Models

The PhD is centred on NLP: benchmark design, retrieval and RAG, model adaptation, training data, expert feedback, and evaluation under ambiguity. Literature provides the demanding test case, but the research questions concern language-model training and evaluation more broadly.

Deadline 8 November 2026 Start 1 February 2027 Location Copenhagen

Read the call and apply

Project summary

Culture is not a collection of isolated facts. It is expressed through context, references, recurring patterns, disagreement, and changing interpretations. CuRe studies whether language models can reason with these forms of meaning—and how their evaluation and training should change when no single answer captures the whole task.

We use Danish literature as a rigorous testbed. Literary interpretation requires models to connect claims to textual evidence, use historical and cultural context, consider plausible alternatives, and respond to criticism. These are difficult capabilities that standard factual and multiple-choice benchmarks do not measure well.

01

Build valid benchmarks

How can we evaluate culturally grounded interpretive reasoning in domains where no unambiguous ground truth exists?

02

Evaluate depth

How can evaluation reflect evidence, ambiguity, disagreement, and context rather than reduce interpretation to a single label?

03

Train better models

How can we adapt and guide LLMs for culturally specific reasoning without reinforcing stereotypes or collapsing plural viewpoints?

Why literature?

Literature is culturally dense and resistant to easy classification. It gives us a controlled but demanding setting for studying capabilities that matter well beyond literary analysis.

Referencing

Recognising quotations, allusions, echoes, and their changing significance across texts and periods.

Reuse

Tracing motifs, clichés, idioms, and recurring cultural patterns without treating culture as a fixed stereotype.

Interpretation

Supporting readings with evidence, comparing alternatives, and preserving meaningful scholarly disagreement.

Research programme

WP1 · Empirical foundations

Data and benchmark development

Curate Danish literary corpora and co-develop interpretive tasks with literary scholars, using evidence-centred benchmark design and annotations that preserve variation.

WP2 · Methodological reflection

Reasoning beyond ground truth

Evaluate retrieval, RAG, and long-context models; study interpretive hallucinations; and develop protocols that separate identifiable errors, evidential support, and plausibility.

WP3 · Model development

Adaptation and training

Test how pretraining mixtures, contextual resources, fine-tuning, and feedback from literary scholars affect cultural reasoning while preserving ambiguity and plural perspectives.

Publications

Perspective preprint2026

From annotation to reasoning: Culture in language models

Daniel Hershcovich, Alexander Conroy, and Jens Bjerring-Hansen

We argue that cultural evaluation must address interpretive depth: whether a model can support a reading with evidence, consider alternatives, and revise it after criticism. The paper connects benchmark design, expert evaluation that preserves disagreement, and experiments on data, context, and feedback.

Read the preprint →

Accepted · CHR 20272027

Retrieving Biblical Intertextual References in Karen Blixen’s Seven Gothic Tales

András Kovacs, Alexander Conroy, Daniel Hershcovich, and Jens Bjerring-Hansen

Accepted at Computational Humanities Research (CHR) 2027 in Manchester for a long oral presentation (22 minutes including Q&A).

We compare sparse and dense retrieval methods for finding transformed cultural references across historical language and translation. Fine-tuning improves retrieval substantially, while expert reassessment shows why benchmark annotations cannot always be treated as exhaustive ground truth.

Read the preprint →

Expected outcomes

BenchmarksEvidence-centred tasks for cultural and literary reasoning

DataOpen corpora and expert annotations grounded in Danish literature

MethodsEvaluation protocols that preserve ambiguity and disagreement

ModelsTraining and adaptation strategies for culturally robust reasoning

Team

Principal Investigator · NLP

Daniel Hershcovich

Department of Computer Science, University of Copenhagen

Co-Principal Investigator · Literary Studies

Jens Bjerring-Hansen

Department of Nordic Studies and Linguistics, University of Copenhagen

Advisor · NLP

Desmond Elliott

Department of Computer Science, University of Copenhagen

Postdoctoral Researcher

Alexander Conroy

Computational literary studies and cultural analytics

Open position · Start February 2027

PhD Fellow

Cultural reasoning, language-model evaluation, and model adaptation

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

Contact

Interested in the project?

For research inquiries and collaboration, contact Daniel Hershcovich at dh@di.ku.dk.

CuRe is funded by the Independent Research Fund Denmark through the thematic programme on artificial intelligence.