Reproduce “Adaptive Localization of Knowledge Negation for Continual LLM Unlearning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Abudukelimu Wuerkaixi, Qizhou Wang, Sen Cui, Wutong Xu, Bo Han, Gang Niu, Masashi Sugiyama, Changshui Zhang
ALKN addresses continual language-model unlearning by adapting parameter updates across successive deletion requests. It fine-tunes the current model on each target, then subtracts a scaled parameter change, using soft labels, a dynamic gradient mask, and lower learning rates for previously activated parameters. On a three-task Harry Potter benchmark, it beats WAGLE on F-Rouge and T-Rouge scores, both proxies rather than proof of deletion.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.