Johannes Bjerva

Research vision

Language as a foundation for trustworthy AI.

My programme connects linguistic theory with controlled experiments on model behaviour. I use semantic analysis and cross-linguistic variation to investigate information leakage, factual reliability, and the limits of current evaluation.

A connected research programme

Research programme

  1. Linguistic foundations

    How do models encode meaning and generalise across languages and cultures?

  2. Trustworthy systems

    How can linguistic insight explain failures in security, privacy, safety, and factuality—and support defensible evaluation?

  3. Next frontier

    How do retrieval, tools, memory, and interaction reshape meaning and reliability?

    Early evidence: tool-elicited reasoning in frontier models · preprint

Established research

Research strands

Discuss a research collaboration →

Research strand

Language-model security and privacy

How language shapes model security and privacy.

My work spans embedding inversion, information leakage, poisoning, and manipulated or machine-generated text across languages.

Core questionHow do linguistic properties shape vulnerabilities, attacks, and defences?

Selected papersChen et al., ACL 2024Yu et al., CCS 2026 · Accepted

Research strand

Factuality and evaluation

Reliable AI requires factual grounding and defensible evaluation.

I study how knowledge grounding, data quality, benchmark design, and language sampling shape claims about model factuality and generalisation.

Core questionHow can outputs be grounded in knowledge, and reliability claims tested beyond narrow benchmarks?

Selected papersLavrinovics et al., ICML 2026Ploeger et al., CL 2026

Research strand

Multilingual NLP

Multilingual does not automatically mean inclusive.

My work combines linguistic typology, careful language sampling, and targeted adaptation, with particular attention to low-resource languages.

Core questionWhat determines whether models and findings transfer across languages?

Applications include explainable educational feedback in Digital Twins for Abundant Feedback, with SEFL exploring how LLM agents can generate feedback data for student assignments.

Selected papersBjerva et al., CL 2019Ren et al., EMNLP 2026

Research strand

Formal and computational semantics

Linguistic structure can make model behaviour easier to analyse.

I draw on formal and computational semantics to analyse model behaviour beyond aggregate performance scores, from compositional meaning representations to discourse context.

The Formal Semantic Methods for AI Safety programme develops this foundation towards explaining language-model behaviour.

Core questionWhich aspects of meaning, context, and inference do current models capture—and which do they miss?

Selected papersAbzianidze et al., EACL 2017Fekete et al., EMNLP 2026

Funded research

Projects and programmes

Selected externally and institutionally funded research, the teams it supports, and work linked to each programme. VBN record counts are snapshots from 5 September 2026.

Single-PI external awards

≈ DKK 25M

Shared or internal programmes

≈ DKK 10M

Full award values; funding shares are shown per project.

Funding

DKK 9.9M

Data Science Investigator (Ascending) · NNF Research Leader Programme

Principal Investigator

Funding share100%

Security & Privacy2025–2030

(LM)²-SEC

Linguistically Motivated Language Model Security

Developing linguistically grounded approaches to security and privacy in language models, including inversion attacks, information leakage, and malicious manipulation across languages.

Funder
Novo Nordisk Foundation
Funded staffing
2 PhDs · 72 PD-months

Linked work

9 VBN records

The first outputs address personal-information memorisation, extraction from diffusion language models, and semantic leakage from image embeddings, including work appearing at ACL and EMNLP 2026 and the diffusion-model memorisation paper accepted at NeurIPS 2026.

Project and linked outputs on VBN

Funding

DKK 6.2M

Sapere Aude: DFF-Research Leader

Principal Investigator

Funding share100%

Security & Privacy2026–2030

TRUST

Building TRUST in Text: Linguistically Motivated Language Model Detection

Investigating whether linguistic signals in generated text can reveal poisoning and support more robust language models.

Funder
Independent Research Fund Denmark · Sapere Aude
Funded staffing
2 PhDs · 12 PD-months

Current status

Started May 2026

The four-year programme began in May 2026. VBN now links the Capable yet Parsimonious preprint, which studies tool-elicited reasoning in frontier models, to the programme.

Project details on VBN

Funding

DKK 2.4M

Technical AI Safety Research

Principal Investigator

Funding share100%

Linguistics & Semantics2026–2028

Formal Semantic Methods for AI Safety

Formal and computational semantics for AI safety

Developing formal and computational semantic methods to analyse and explain language-model behaviour in multilingual settings.

Funder
Coefficient Giving
Funded staffing
36 PD-months

Current status

Started May 2026

The two-year programme began in May 2026. Early linked work includes Capable yet Parsimonious, a preprint analysing the structure and efficiency of tool-elicited reasoning.

Project details on VBN

Funding

≈ DKK 7M

AI:X Labs · Aalborg University

Lab Director / Principal Investigator · Co-led with Qiongxiu Li

Funding share50% · co-led

Security & Privacy2025–2029

AI:SECURITY

AI-enabled threats and secure AI systems

Combining NLP and cybersecurity research to study AI-enabled threats and the security of language models in societal applications.

Funder
AI:X Labs · Aalborg University
Funded staffing
4 PhDs · main supervisor for 2

Linked work

7 VBN records

Linked work spans multilingual memorisation, personal-information leakage, privacy-preserving graph aggregation, and safer model sharing, with papers at EMNLP 2025, ACL 2026, and ICASSP 2026.

Project and linked outputs on VBN

Funding

DKK 1.1M

Industrial PhD · Industrial Researcher Programme

Principal Investigator

Factuality & Evaluation2024–2027

Guarantees of Factuality in LLM-based Extraction of Financial KPIs

Industrial PhD in collaboration with ALIPES ApS

Research on factual, knowledge-grounded extraction of financial indicators from earnings transcripts and company filings.

Funder
Innovation Fund Denmark
Funded staffing
Industrial PhD

Linked work

1 dataset · 2 published papers

The project has delivered HiFi-KPI, a public dataset with an LREC 2026 paper, and an ACL 2026 Industry Track paper on KPI extraction from earnings calls.

Project and linked outputs on VBN

Funding

DKK 5M

Semper Ardens: Accelerate · Carlsberg Foundation

Principal Investigator

Funding share100%

Multilingual NLP2022–2026

Multilingual Modelling for Resource-Poor Languages

Semper Ardens: Accelerate

Fundamental research on linguistic typology, multilingual modelling, and evaluation for languages underserved by current NLP systems.

Funder
Carlsberg Foundation
Funded staffing
3 PhDs · 36 PD-months

Related award

Google Award for Inclusion Research · CREOLE

DKK 425k · 6 PD-months

Funding share100%

Linked work

44 VBN records · 35 activities

Completed in August 2026. Linked research spans typological diversity, multilingual embedding inversion, and low-resource evaluation; EMNLP 2026 papers examine semantic recovery from image embeddings and information structure in vision-language models.

Project and linked outputs on VBN

Funding

DKK 3M

Villum Synergy · Villum Foundation

Principal Investigator / NLP methods lead · Co-led with Euan Lindsay

Funding share50% · co-led

NLP for Education2024–2026

Digital Twins for Abundant Feedback

Novel Feedback Paradigms via Explainable Multilingual NLP

Studying explainable multilingual NLP for scalable, high-quality feedback in education.

Funder
Villum Foundation · Synergy
Funded staffing
24 PD-months

Linked work

16 VBN records · 7 media items

Active, with a scheduled end in December 2026. Linked work connects synthetic educational feedback with LLM agents and knowledge-grounded factuality, alongside public discussion of AI-supported teaching and assessment.

Project and linked outputs on VBN

Staffing figures describe funded positions; “PD-months” means funded postdoctoral researcher months. VBN record counts may include separate preprint and published versions of the same work. Outputs can be linked to multiple programmes: VBN associates Few-shot Semantic Recovery Attacks on Image Embeddings with (LM)²-SEC, AI:SECURITY, and the Carlsberg programme.

Coefficient Giving logo: Aarongertler, CC BY 4.0. Resized; displayed with inverted colours on dark backgrounds.