Mitigating Forgetting in Continually Pretraining MoE-LLMs by Adding and Chilling Experts

Raymond Li, Charles Guille-Escuret, Denis Kocetkov, Joel Lamy-Poirier, Luke Kumar, Max Tian, Sebastien Paquet, Christopher Pal, Torsten Scholak

Keywords: LLM, Continual Learning, Lifelong Learning, Mixture of Experts, Pretraining

Abstract: As model training requires more and more compute, the cost of re-training models to support new data or domains increases as well. Methods to adapt existing models to new data distributions are crucial to avoid spending redundant compute re-training models from scratch. However, naive finetuning often incurs forgetting of previously learned capabilities. In this paper, we analyse how different factors such as model size, dataset size and replay data impact forgetting when adapting models to new data distributions. We also propose to increase the capacity of Mixture-of-experts models by adding new experts and reducing the learning rate of the old model weights. Our experiments show that this simple method allows to reduce forgetting and learn efficiently on the new domain.

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.

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: 10647