EPI: Dynamic Parameter Isolation Framework Solves Catastrophic Forgetting in Large Model Fine-Tuning

This article proposes the Evolving Parameter Isolation (EPI) framework, which addresses task interference and catastrophic forgetting in supervised fine-tuning by dynamically updating parameter isolation masks. Based on the key finding that parameter importance drifts over the training process, this framework breaks through the limitations of static parameter isolation methods and achieves a balance between knowledge retention and new knowledge learning.

Supervised Fine-Tuning (SFT) of large language models is a key step to adapt models to specific tasks, but it faces two major challenges: 1. Task interference: Parameter updates between different tasks conflict with each other; 2. Catastrophic forgetting: Forgetting knowledge of old tasks when learning new ones.

Recent studies mitigate the problem by isolating task-critical parameters, with the basic idea of identifying important parameters for specific tasks and freezing them to protect existing knowledge. However, these methods have assumption flaws: they assume that parameter importance remains unchanged once determined, using static solutions to solve dynamic problems.

Empirical studies have found that parameter importance exhibits temporal drift: some parameters are important for Task A in the early stages of training but become secondary later, while others are critical for Task B in the later stages; fixed isolation masks cannot adapt to such dynamic changes, challenging the basic assumptions of static methods.

The EPI framework achieves dynamic adjustment through the following core mechanisms: 1. Online importance estimation: Continuously monitor the gradient signals of each parameter for the current task; 2. Periodic mask update: Regularly update the isolation mask instead of keeping it fixed; 3. Dynamic protection strategy: Protect newly emerging task-critical parameters and release outdated parameters to restore plasticity. Compared to static methods, EPI can adapt to training dynamics, enhance plasticity, and reduce interference.

In experiments on diverse multi-task benchmarks, EPI consistently reduces task interference and catastrophic forgetting, outperforming static isolation and standard fine-tuning, and improving overall generalization ability.

The study emphasizes the need for isolation mechanisms to synchronize with learning dynamics: Model learning is a dynamic process, and fixed strategies are difficult to adapt; parameter importance is training-dependent and requires continuous evaluation; dynamic isolation strategies can better balance knowledge retention and new knowledge learning.

EPI reveals that parameter importance is not a static attribute but a dynamic feature that evolves with training, a finding that may influence the development of other model adaptation technologies. Future research directions include: exploring more efficient online importance estimation methods, studying the evolution patterns of parameter importance under different architectures and task types, and extending the dynamic isolation idea to other model adaptation scenarios.