Effective LLM Knowledge Learning Requires Rethinking Generalization
Keywords: knowledge learning, generalization, large language models, knowledge acquisition
Abstract: Large language models (LLMs) are trained on a substantial amount of documents that contain extensive world knowledge. However, it is still not well-understood how knowledge is acquired via autoregressive pre-training and extracted via question-answering. This lack of understanding greatly hinders effective knowledge learning, especially for continued pre-training on up-to-date information, as this evolving information often does not have diverse repetitions like foundational knowledge. In this paper, we focus on understanding and improving LLM knowledge learning. We found and verified that knowledge learning for LLMs can be deemed as an implicit supervised task hidden in the autoregressive pre-training objective. Our findings suggest that knowledge learning for LLMs would benefit from methods designed to improve generalization ability for supervised tasks. Based on our analysis, we propose to diversify training documents’ formats as data augmentation to grow in-distribution samples. This data augmentation method does not present the risk of altering the facts embedded in documents as text paraphrasing. We also introduce sharpness-aware minimization as an effective optimization algorithm to better improve generalization. Moreover, we adapt our method to instruction tuning for generalization to various phrasings of questions. Extensive experiment results validate our findings and demonstrate our methods’ effectiveness in improving knowledge learning in both the continued pre-training and instruction tuning stages. This paper offers new perspectives and insights to interpret and design effective strategies for LLM knowledge learning.
Primary Area: other topics in machine learning (i.e., none of the above)
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: 10417
Loading