LiteLoRA Reduces Adapters in Continual Fine-Tuning by Reusing Representations.

Low-Rank Adaptation (LoRA) has become a standard, parameter-efficient method for fine-tuning large pre-trained models, especially in continual learning (CL) scenarios where models adapt to new tasks sequentially. The conventional approach assumes that each new task necessitates its own dedicated low-rank adapter. However, new research empirically and structurally challenges this assumption, revealing substantial low-rank redundancy among task-specific LoRA adapters. The study demonstrates that the subspaces spanned by adapters trained on different tasks often overlap significantly, implying that earlier adapters can effectively represent later tasks. Building on this insight, researchers introduce LiteLoRA, a novel plug-and-play gating mechanism. This mechanism intelligently learns during training whether to recruit a new adapter for a task or to reuse existing low-rank representations. LiteLoRA has been shown to reduce the number of active adapters by 20-70% while matching or even exceeding state-of-the-art performance on standard CL benchmarks. This finding highlights the pervasive nature of structural redundancy in continual fine-tuning and confirms that selective learning is sufficient to achieve model stability without sacrificing plasticity, leading to more efficient and scalable continual learning systems.