Reproduce “RADNPO: Reference-free Adaptive Negative Preference Optimization for 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
Shenghan Tan, Ziyi Zhou, Wenpeng Hu, Mengyuan Zhang
RADNPO performs approximate machine unlearning by comparing each forget-token probability with high-probability alternatives from the model’s current next-token distribution, then scaling a soft-clamped contrastive objective using confidence and partial entropy. On TOFU Forget10 with LLaMA2-7B-chat, it reports EM 0.382, ES 0.040, and MU 0.704 versus NPO. Re-exposure to the exact forget set raises EM to 0.968, so deletion is not guaranteed.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.