KLoB: a Benchmark for Assessing Knowledge Localization Methods in Language Models

Yiming Ju, Huimin Ma, Xingrun Xing, Zhixiong Zeng

Abstract: Recently, the Locate-Then-Edit paradigm has emerged as one of the main approaches in changing factual knowledge stored in the Language models. However, there is a lack of research on whether present localization methods can pinpoint the exact parameters embedding the desired knowledge. Moreover, although many researchers have questioned the validity of locality hypothesis of factual knowledge, no method is provided to test the hypothesis for more in-depth discussion and research. Therefore, we introduce KLoB, a benchmark examining three essential properties that a reliable knowledge localization method should satisfy. We evaluate several knowledge localization methods from existing knowledge editing approaches on the KLoB benchmark. The results indicate that while these methods outperform random selection, their overall performance remains suboptimal. KLoB can serve as a benchmark for evaluating existing localization methods in language models, and it contributes a method to reassessing the validity of the locality hypothesis of factual knowledge. KLoB is publicly available at an anonymous GitHub repository: https://github.com/anon6662/KLoB.

External IDs:dblp:conf/pricai/JuMXZ24