Towards Domain-Centric Artificial Intelligence: Bridging General Capabilities and Domain-Specific Context

Licentiate thesis, 2026

Foundation models and Large Language Models (LLMs) have strong general capabilities. They can understand language, reason across different tasks, generate code, and solve problems in areas where they were not specifically trained. This broad capability makes them powerful starting points for real-world AI systems. However, for high-stakes domains, such as automotive software engineering, AI systems must do more than provide plausible answers. They must follow domain rules, respect data structures, handle operational constraints, and produce reasoning that experts can check and trust. This creates a gap between general capability and domain-specific reliability.

This thesis argues for Domain-Centric AI: the design of AI systems that are generalizable across domains, adaptable to target domains, and able to reason reliably within specific operational domains. These levels build on one another. Generalization provides the model-level foundation. Adaptation aligns this foundation with a target domain. Domain-specific system design then enforces the rules, workflows, and constraints needed for reliable use.

The thesis explores this progression through four papers. The first paper surveys meta-learning methods for domain generalization and shows how models can become more robust to unseen domains. The second paper extends this perspective to vision-language models by introducing latent domain prompt learning for domain generalization. The third and fourth papers focus on industrial LLM-based agent systems for automotive software release analytics. They demonstrate how general LLM capabilities can be embedded in multi-agent and pipeline designs to support informed decision-making.

Together, the studies show reliable AI in domain-specific settings can be designed by combining a flexible model core with a constraining system around it. The core model must generalize across unseen domains. The surrounding system must enforce domain logic. By bridging general AI capabilities with structured domain-specific context, Domain-Centric AI can improve robustness, reduce manual effort, and support more reliable decision-making in safety-critical workflows.