Reproduce “Auditable Unlearning for LLMs” 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
Piyush Tiwary, Anushka Srivastava, Prathosh A.P
GapWiden and GapVeil create provider-side separation between forget and retain examples, then screen public drafts using that signal. On TOFU forget10 with Llama 3 8B, averages across 11 methods change from 0.904 to 0.927 for forgetting, 0.631 to 0.955 for retention, 0.753 to 0.993 for leakage robustness, and 0.944 to 0.997 for auditability. These benchmark results do not prove universal deletion or security.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.