Reproduce “Preserving Utility in Continual LLM Unlearning via Orthogonal Projections” 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
Juan Belieni, Ana Carolina Erthal, Eliezer de Souza da Silva, Diego Mesquita
The authors introduce COP, which stores compact activation bases from earlier forget sets and projects later updates away from those directions. On TOFU after five requests, NPO+COP yields MU 0.532 versus 0.496 without COP, with FR, ES, and EM of 0.30, 0.05, and 0.56, versus 0.32, 0.08, and 0.63. Protection is approximate, not certified deletion.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.