CleanEdit
Retention-Aware Pruning & Bounded Replay for Lifelong Model Editing
Haoyuan Song, Haihua Luo, Ming Wang, Yisu Wang, Qi Xu, MingKun Xu, Wenjie Huang, Xuming Ran
October 2025
Abstract
Lifelong model editing often suffers from memory pollution — where redundant or conflicting edits accumulate, leading to instability, bloated inference costs, and loss of reliability. We introduce CleanEdit, a novel framework that addresses these challenges through Retention-Aware Pruning and Bounded Replay mechanisms. Our approach statistically detects and removes destabilizing edits while recycling supervision signals from pruned edits. We propose three scheduling modes — Comprehensive, Progressive, and Dynamic — that adapt to different deployment needs. Extensive experiments on SCOTUS (classification) and zsRE (QA) benchmarks demonstrate over 10% improvements in balancing knowledge retention and new edit fidelity. CleanEdit transforms edit memory into a self-maintaining system, ensuring stability and reliability for long-term deployments.
Cite This Work
@article{cleanedit2025,
title={CleanEdit: Retention-Aware Pruning & Bounded Replay for Lifelong Model Editing},
author={Haoyuan Song and Haihua Luo and Ming Wang and Yisu Wang and Qi Xu and MingKun Xu and Wenjie Huang and Xuming Ran},
archivePrefix={arXiv},
year={2025}
}