Combating the Generalization-Forgetting Trade-off in Continual Learning: A Cautious Passive Low-Rank Approach

ICLR 2025 Conference Submission12732 Authors

28 Sept 2024 (modified: 13 Oct 2024)ICLR 2025 Conference SubmissionEveryoneRevisionsBibTeXCC BY 4.0
Keywords: continual learning, LLMs
TL;DR: We propose a parameter-efficient approach for continual learning in LLMs
Abstract: Large Language Models (LLMs) have shown remarkable capabilities through wide-scale pre-training on a wide range of domains. However, they often suffer from catastrophic forgetting when learning sequential tasks. In this paper, we propose a novel parameter-efficient approach for continual learning in LLMs, which empirically explores the role of different effective layerwise ranks, leveraging lower ranks to mitigate catastrophic forgetting of previous tasks and higher ranks to enhance generalization on new tasks. By employing a subspace similarity metric that evaluates the orthogonality of low-rank subspaces between tasks, we gradually increase the rank of layerwise matrices for each new task, minimizing interference with previously learned tasks while enhancing generalization. Experimental results on standard continual learning benchmarks and challenging math benchmarks demonstrate that our method outperforms existing state-of-the-art approaches, effectively mitigating forgetting, improving task performance, and maintaining strong generalization to unseen tasks in a memory-efficient manner.
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: 12732
Loading