Unraveling LoRA Interference: Orthogonal Subspaces for Robust Model Merging
This addresses a bottleneck in deploying and storing multiple task-specific models for practitioners, offering a plug-and-play solution for merging LoRA models, though it is incremental as it builds on existing merging methods.
The paper tackles the problem of performance degradation when merging models fine-tuned with LoRA, showing that it arises from an interplay between model parameters and data distributions, and proposes OSRM to constrain LoRA subspaces prior to fine-tuning, which boosts merging performance and preserves single-task accuracy across eight datasets and multiple LMs.
Fine-tuning large language models (LMs) for individual tasks yields strong performance but is expensive for deployment and storage. Recent works explore model merging to combine multiple task-specific models into a single multi-task model without additional training. However, existing merging methods often fail for models fine-tuned with low-rank adaptation (LoRA), due to significant performance degradation. In this paper, we show that this issue arises from a previously overlooked interplay between model parameters and data distributions. We propose Orthogonal Subspaces for Robust model Merging (OSRM) to constrain the LoRA subspace *prior* to fine-tuning, ensuring that updates relevant to one task do not adversely shift outputs for others. Our approach can seamlessly integrate with most existing merging algorithms, reducing the unintended interference among tasks. Extensive experiments on eight datasets, tested with three widely used LMs and two large LMs, demonstrate that our method not only boosts merging performance but also preserves single-task accuracy. Furthermore, our approach exhibits greater robustness to the hyperparameters of merging. These results highlight the importance of data-parameter interaction in model merging and offer a plug-and-play solution for merging LoRA models.