Title:	Learning Adaptive Digital Representations for Structured data and Graphs at Scale
DNr:	Berzelius-2026-249
Project Type:	LiU Berzelius
Principal Investigator:	Sarunas Girdzijauskas <sarunasg@kth.se>
Affiliation:	Kungliga Tekniska högskolan
Duration:	2026-10-02 – 2027-05-01
Classification:	10210
Keywords:

Abstract

Recent advances in foundation models have transformed artificial intelligence in domains such as language and vision. However, much of the world's information exists in structured forms, including tables, graphs, relational databases, network telemetry, and scientific data. Developing scalable methods that can learn transferable representations from such data remains a fundamental challenge. This project aims to advance representation learning for structured data and graphs through self-supervised learning, adaptive modeling, knowledge transfer, model compression, and efficient inference techniques, organized around six interconnected research objectives. Particular emphasis will be placed on developing computationally efficient and deployable AI systems, including quantized and compressed model representations and KV-cache-aware inference strategies. The project is expected to produce scientific publications, open-source foundation models for tabular data, and efficient methodologies for scalable AI on structured domains. Access to Berzelius will enable large-scale experimentation across diverse datasets and graph domains, supporting the development of adaptive, scalable, and sustainable AI. Therefore, we request 4000 gpu-hours/month for six months in Berzelius Hopper.