A3 Lab is a research group at Aarhus University advancing adaptive and agentic AI for reliable deployment under non-stationarity. We develop multimodal foundation models and test-time learning methods that can quantify uncertainty, and adapt safely on the fly—minimizing costly re-training while maintaining robustness, auditability, and governance in real-world, safety-critical settings.
We study multimodal foundation models (vision-language and beyond) for robust perception and decision making under distribution shift. Our work includes prompting and instruction following, interactive and agentic systems, and explainability and faithful reasoning so that outputs are not only accurate but also verifiable.
We develop self-supervised learning methods to build stronger representations that transfer across tasks and domains. We also explore synthetic data and generative modeling to improve robustness, cover rare conditions, and reduce dependence on labeled data.
We design lightweight methods that update models at inference time using unlabeled or weak supervision. Key topics include domain generalization / test-time adaptation, parameter-efficient continual updates, and unknown-class / novelty recognition to handle evolving and recurring domain shifts.
We investigate multi-agent coordination and planning for complex AI pipelines, including human–AI teaming and decision support. To enable practical deployment, we work on efficient model design (distillation, efficient transformers, TinyML-style constraints) and rigorous evaluation protocols for reliability under drift.
A3 Lab contributes to the research community through workshops, reviewing, and open dissemination of research outcomes.
