Title:	Foundation model adaptation
DNr:	Berzelius-2026-183
Project Type:	LiU Berzelius
Principal Investigator:	Kevin Smith <ksmith@kth.se>
Affiliation:	Kungliga Tekniska högskolan
Duration:	2026-06-26 – 2027-01-01
Classification:	20208
Keywords:

Abstract

Foundation models have become powerful general-purpose tools, but adapting them to new domains and modalities remains challenging, particularly when labeled data are scarce. In this continuation project, we will investigate methods for multimodal transfer learning and adaptation: how large pretrained models can be adapted, combined, and aligned to encode information across heterogeneous data sources and tasks. This is a broad and active research direction, and we anticipate exploring it from several angles. Examples of directions we expect to pursue include lightweight alignment of representations across different datasets, domains, or model families without paired supervision, and the use of multimodal foundation models to integrate heterogeneous clinical signals for downstream tasks such as patient trajectory prediction. More generally, the unifying theme is the adaptation of multimodal foundation models to new modalities, tasks, and domains, a setting that arises across a wide range of real-world applications.