Title: Enhancing Efficiency and Collaboration in Fine-Tuning Through Task Similarity

Abstract: The ability to adapt to unseen downstream applications is a defining characteristic of foundation models. Techniques like LoRA have emerged as prominent parameter-efficient fine-tuning strategies, allowing these large-scale models to adapt effectively using limited, high-quality, labeled data. To address the challenge of data scarcity during the fine-tuning process, this study proposes leveraging the inherent similarities between tasks across various downstream users. The underlying premise is that users dealing with analogous tasks can mutually enhance one another’s effective training data volume.

We introduce Collaborative Low-Rank Adaptation (CoLoRA), a method designed to fine-tune personalized foundation models by exploiting task similarities in a collaborative and efficient manner. CoLoRA operates on the principle of training two types of components: a shared adapter that encapsulates the commonalities across all tasks, and personalized adapters customized for specific user tasks. We provide a theoretical analysis of CoLoRA within the context of heterogeneous linear regression, offering provable guarantees for the recovery of ground truth. Furthermore, our natural language processing experiments, conducted across varying degrees of task similarity, confirm that joint training with similar tasks leads to significant performance improvements for individual models.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC