How new data pollutes LLM knowledge and how to dilute it
Keywords: fine-tuning, hallucinations, knowledge injection, memory, LLMs
Abstract: Understanding how the learning of new texts alter the existing knowledge in a large language model is of great importance, because it is through these accumulated changes that the LLM was initially pre-trained, and is also through such changes that continual, new learning in LLMs can proceed. As a result, both desirable alterations (i.e. generalization) and undesirable alterations (i.e. hallucination) can occur. Here, we study the learning of new texts, one at a time, and ask: how does it impact the underlying LLM knowledge?
We show that learning new texts induce 'priming', an undesirable effect that pollutes existing knowledge where it should not.
Centrally, we demonstrate that we can predict how much priming will happen after learning, using token probability before learning. This was empirically robust across models (PALM-2-xs/s, Gemma-2b, Llama-2-7b), of various sizes, and training stages. To show this, we created a new dataset, called "Outlandish" consisting of 1320 different samples with diverse textual characteristics. Finally, we propose two strategies to mitigate the spread of priming: first, a simple text augmentation technique which we call the "stepping-stone'', and second, a novel update pruning technique ("ignore-k"). These decrease priming by a median of 50%-75% and 50%-95% respectively depending on the model architecture, and enhance the specificity of new learning in language models. The dataset and reproducible findings can be found [LINK omitted for double blind review].
Primary Area: interpretability and explainable AI
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: 13888
Loading