MQuAKE-Remastered: Multi-Hop Knowledge Editing Can Only Be Advanced With Reliable Evaluations

Shaochen Zhong, Yifan Lu, Lize Shao, Bhargav Bhushanam, Xiaocong Du, Louis Feng, Yixin Wan, Yucheng Shi, Daochen Zha, Yiwei Wang, Ninghao Liu, Kaixiong Zhou, Shuai Xu, Vipin Chaudhary, Xia Hu

Keywords: knowledge edit, model edit, multi-hop, question answering, natural language processing, dataset audit

TL;DR: Updating one knowledge fact will produce a ripple effect, making multi-hop knowledge editing (MHKE) a desired capability for reliable LLMs. We reveal many unknown errors of MQuAKE — the most popular MHKE dataset — though an audit and fix everything.

Abstract:

Large language models (LLMs) can give out erroneous answers to factually rooted questions either as a result of undesired training outcomes or simply because the world has moved on after a certain knowledge cutoff date. Under such scenarios, *knowledge editing* often comes to the rescue by delivering efficient patches for such erroneous answers without significantly altering the rests, where many editing methods have seen reasonable success when the editing targets are simple and direct (e.g., *"what club does Lionel Messi currently play for?"*).

However, knowledge fragments like this are often deeply intertwined in the real world, making effectively propagating the editing effect to non-directly related questions a practical challenge (e.g., *"who is the offspring of the owner of the club that Messi currently plays for?"*). Prior arts have coined this task as *multi-hop knowledge editing* with the most popular dataset being MQuAKE, serving as the sole evaluation benchmark for many later proposed editing methods due to the expensive nature of making knowledge editing datasets at scale.

In this work, we reveal that **up to 33\% or 76\% of MQuAKE's questions and ground truth labels are, in fact, corrupted in various fashions due to some unintentional clerical or procedural oversights**. Our work provides a detailed audit of MQuAKE's error pattern and a comprehensive fix without sacrificing its dataset capacity. Additionally, we benchmarked almost all proposed \mquake{}-evaluated editing methods on our post-fix dataset, \mquaker{}. It is our observation that many methods try to overfit the original \mquake{} by exploiting some data-specific properties of \mquake{}. We provide a guideline on how to faithfully approach such datasets and show that a simple, minimally invasive approach can bring excellent editing performance without such exploitation. Please refer to [`https://github.com/henryzhongsc/MQuAKE-Remastered`](https://github.com/henryzhongsc/MQuAKE-Remastered) and supplemental material for assets.

Submission Number: 840