Pioneering Applied Artificial Intelligence
From foundational NLP research to scalable medical data systems. We transform unstructured narratives into actionable science.
RESEARCH AREAS
NLP and Deep Learning
Basic Research in NLP
Our basic research focuses on foundation models for NLP, particularly reasoning, adaptation, and evaluation. We investigate how these models represent knowledge, reason over information, and generalize across tasks and domains. We designed and trained foundation models, such as Soofi, MedAlpaca, and medBert.
View Publications in Basic Research in NLPNLP in Healthcare
Together with leading health care providers in Germany we focuses on language models for representing patients and predicting clinical outcomes. More recently, we investigate long-context models and medical reasoning with foundation models.
View Publications in NLP in HealthcareExplaining and Benchmarking
We design methods to explain and benchmark foundation models, from visualizing internal representations to evaluating long-context understanding and reasoning. We study their capabilities, failure modes, and reliability in realistic settings.
View Publications in Explaining and BenchmarkingApplications and Demos
We build interactive systems that make foundation models accessible and explainable. Our demos visualize model representations, support human-in-the-loop learning, and interactively explain model predictions.
View Publications in Applications and DemosRESEARCH SCENARIOS
Two research streams
Modular and Self-Improving Foundation Model Systems
Our researches modular foundation models that combine specialized components instead of relying solely on monolithic LLMs.
KEY HIGHLIGHTS
• SOOFI: A sovereign German-English foundation model based on a hybrid MoE Mamba-Transformer architecture.
• Reasoning Post-Training: Improving capabilities after pretraining.
• Global Exchange: Co-organization of the Shonan Seminar for empowering local open-source LLMs.
Interpretable Patient Representations with Foundation Models
Our develops interpretable patient representations from clinical text to support diagnosis, outcome prediction, and clinical decision support.
KEY HIGHLIGHTS
• medBERT.de: Domain-specific medical language models.
• S-PROTO & SPONGE: Sparse prototypical representations and cross-lingual knowledge transfer.
• Clinical Evaluation: CliniBench & DeepICD-R1 for medical reasoning and rare clinical outcomes.
RECENT ACTIVE PROJECTS
Ongoing work at DATEXIS
Appl-FM — Berlin Initiative for Applied Foundation Model Research
The Berlin Initiative for Applied Foundation Model Research (Appl-FM) is a research impulse funded by the German Research Foundation (DFG) at the Berliner Hochschule für…
Read More about Appl-FM — Berlin Initiative for Applied Foundation Model ResearchAI4SCM — Machine Reading in Supply Chain Management für KMUs
DATEXIS investigates data-efficient machine reading and representation learning for detecting and classifying rare events in supply-chain data.
Read More about AI4SCM — Machine Reading in Supply Chain Management für KMUsSWAMBIT — Shared Workspace And Memory for Brain-Inspired Transformers
DATEXIS investigates brain-inspired modular AI architectures combining specialist transformer modules, shared workspace and memory, sparsity, and global feedback mechanisms.
Read More about SWAMBIT — Shared Workspace And Memory for Brain-Inspired TransformersSOOFI. A Sovereign, Open-Source Foundation Model.
Soofi is a research project for sovereign open-source foundation models from Germany, launched in September 2025. Leading research institutions, universities, and AI companies are collaborating…
Read More about SOOFI. A Sovereign, Open-Source Foundation Model.COMFORT — Computational Models FOR patienT stratification in urologic cancers
COMFORT is a Horizon Europe research project developing trustworthy, multimodal AI models to support the diagnosis and treatment of prostate and kidney cancer, launched in…
Read More about COMFORT — Computational Models FOR patienT stratification in urologic cancersOUR PARTNERS
Partners across our projects
- Zalando
- Siemens
- Exasol
- ellamind
- Merantix Momentum
- Merantix Labs
- Quantib
- Phönix-PACS
- 2txt NLG
- Neofonie
- Ubermetrics
LAB ASSETS
Models, demos, and benchmarks
Medical-DDX
Medical-DDX
Medical-Negation
Medical-DDX
Explanibale AI
Medical-Knowledge Retrieval
Medical-Knowledge Retrieval
Named Entity Linkage
PROJECT ARCHIVE