Reproduce “MuEdit: An efficient multi-task editing method towards inter-domain knowledge conflicts” 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

Songshi Liang, Ting-En Lin, Zixuan Huang, Yuchuan Wu, Yongbin Li, Zihe Wang, Rui Yan

Mu-Edit addresses sequential model editing across functional domains by estimating conflicts between tasks’ preservation constraints, ordering tasks by adjacent pairwise conflict, and selectively relaxing constraints through low-rank compression. On Llama3-8B under Best Order, it reaches 0.8845 Specificity on ZsRE, versus 0.7098 for AnyEdit under the same ordering. This measures desired-target accuracy on edited examples, not cumulative forgetting.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.