I-Lora: Iterative Merging of Routing-Tuned Low-Rank Adapters for Multi-task Learning
Keywords: Multitask learning, Low-rank adaption, Vision-language-models
Abstract: The advancement of vision-language models has significantly boosted the performance of embodied and game AI, endowing them with more robust general visual understanding capabilities and logical abilities for action planning. However, the substantial computational cost of model training and the performance degradation during fine-tuning limit the models' ability to learn emerging new tasks continually. Creating a versatile and dynamically updatable vision-language model is an essential area of research. To this end, we propose a Low-Rank Adapter-based fine-tuning approach called I-LoRA, which enables iterative and independent learning of new tasks while preserving the logical capabilities of the previously trained model. Specifically, we first design the routing-tuning method to minimize the impact of original capabilities from the new task by minimizing activation values of LoRA matrices as low as possible in the general task. Secondly, we propose a novel approach to iteratively merge new adapters, allowing for continuous integration of adapters trained on new tasks without being influenced by task order, thereby reducing interference between them. Finally, we conducted extensive experiments on public datasets with significant behavioral and logical differences between tasks. The results demonstrate that our approach achieves excellent single-task performance, strong multi-task compatibility, and flexible scalability without increasing the number of model parameters.
Primary Area: transfer learning, meta learning, and lifelong learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 10699
Loading