Intern
    Data Science Chair

    DEMAnD-LM – Domain-Enhanced Multi-Agent and Mixture-of-Experts Language Models

    The project Domain-Enhanced Multi-Agent and Mixture-of-Experts Language Models (DEMAnD-LM) investigates novel architectures for integrating structured domain knowledge into generative AI systems. The project develops methods that allow large language models to incorporate heterogeneous knowledge sources and to specialize through modular expert components. By combining language models with structured knowledge bases and expert models, DEMAnD-LM aims to enable reliable AI systems that can answer complex technical questions and support knowledge-intensive workflows while maintaining efficiency and strong data protection guarantees.


    DEMAnD-LM is a joint research endeavor between vAudience GmbH, denkbares GmbH, and the University of Würzburg.

    Within the project, the university contributes research on domain knowledge integration and domain-specific applications, including the cyber security domain.


    The objectives of the project can be summarized as follows:

     

    • Development of methods for the continuous integration of heterogeneous domain knowledge, for example through knowledge graphs and parameter-efficient architectures for incorporating structured knowledge into language models.
    • Development of domain-specific mixture-of-experts architectures that enable the specialization of language models on distinct technical knowledge areas.
    • Design of intelligent routing mechanisms that dynamically select between expert models, assistance models, and structured data sources depending on the user request.
    • Development of multi-agent orchestration strategies for complex tasks. In the cyber security domain, this includes the coordination of specialized agents that utilize cyber security knowledge and expert models to support activities such as vulnerability analysis and penetration testing across different operational phases.

    Contact

    If you are interested in learning more about the project, collaborating with us, or conducting a thesis or student internship in this area, please feel free to contact Daniel Schlör.