Dissertation: Efficient and Controllable Foundation Model Finetuning for Continual Learning

What did you study?

Foundation models, such as large language models like GPT, have achieved remarkable performance across a wide range of tasks. However, while the world is dynamically changing, the knowledge learned by a model is largely fixed after training. Foundation models therefore need to adapt to new tasks while retaining knowledge from previous tasks, which is known as continual learning. The high computational cost of retraining foundation models and the problem of catastrophic forgetting make continual learning a major challenge.

I studied how foundation models can efficiently learn new task knowledge without forgetting previously learned knowledge. My research focused on the effect of model finetuning in the continual adaptation of foundation models, specifically on how a model can be adapted by updating only a small number of lightweight modules without retraining foundation model. The target is to make this adaptation process more efficient, scalable and controllable. 

What were the results of your study?

The research shows that foundation models can be efficiently adapted to new tasks without retraining the foundation model. In the dissertation, I established a unified continual adaptation framework consisting of a frozen pre-trained backbone and lightweight trainable modules. The proposed methods demonstrated powerful performance across extensive benchmarks, enabling foundation models to learn new tasks efficiently while retaining previously acquired knowledge. The research also indicates that increasing the number of adapter connections does not necessarily improve model performance; removing the key-value mechanism can reduce interference between tasks, and representation finetuning can adapt models through an optimization process with explicit objective. In addition, semantic routing can enable reversible model editing, allowing specific edits to be precisely reversed. 

How can the results be applied? 

These results can be applied to AI systems that need to be employed in dynamic environments for a long time. When new tasks come, models can adapt to them by introducing lightweight trainable modules rather than retraining the foundation model. This can reduce computational costs and make AI systems easier to maintain when facing dynamic environments. For large language models, semantic routing can also enable reversible editing, allowing precise revision for specific knowledge edits. 

What new insights did the research contribute to the topic? 

This dissertation establishes a unified framework for continual adaptation and proposes a paradigm for continual finetuning of foundation models. The framework consists of a frozen pre-trained backbone and lightweight trainable modules, and investigates model adaptation at different levels, ranging from the parameter level and representation level to the knowledge level. The research shows that foundation models can be efficiently adapted to new knowledge by keeping the pre-trained backbone frozen and introducing lightweight trainable modules.

M.Sc. Haihua Luo defends their doctoral dissertation, "Efficient and Controllable Foundation Model Finetuning for Continual Learning" on September 11, 2026 at 12:00 in hall S212 (Vanha Juhlasali). Opponent is Professor Jussi Tohka from the University of Eastern Finland, and Custos is Professor Tommi Kärkkäinen from the University of Jyväskylä.

The language of the dissertation and the event is English. The event can also be followed via Zoom (Meeting ID:  683 4564 6118).