Unlearning in large language models aims to remove specific knowledge — for privacy or safety — while preserving general capabilities. The typical approach treats the model as a static artifact: erase once, then ship. But that assumption breaks for models that receive updates after release.
The new arXiv paper, titled "Not All Change Is Damage: Update-Matched Unlearning for Large Language Models," argues that this static goal is insufficient when a model continues to be fine-tuned post-release. Any new fine-tuning can reintroduce or alter knowledge that was supposed to be removed, or create new risks that were not present at the original unlearning step.
To address this, the authors propose an approach they call "update-matched unlearning." While the abstract does not detail the method, the name suggests that unlearning should be aligned with the model's update schedule — so that removals stay effective as the model evolves. The paper highlights a gap in current unlearning practice and offers a direction for making it robust to ongoing model changes.