Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Yiming Ju, Xingrun Xing, Zhixiong Zeng

TL;DR

The paper tackles the lack of rigorous benchmarks for evaluating knowledge locating methods in language models and the ongoing debate about the locality of factual knowledge within model parameters. It introduces KLoB, a benchmark that enforces three properties—Consistency, Relevance, and Unbiasedness—to assess locating results, and provides three subtasks (KLoB-c, KLoB-r, KLoB-u) anchored in Wikidata and MQUAKE data. KLoB also defines relative evaluation metrics, RSim and RSD, to enable fair cross-method comparisons across varying localization scopes. By offering data construction protocols and evaluation tools, the work supplies a concrete platform to benchmark locating methods and to reassess the locality hypothesis, with a public GitHub release for reproducibility and extension.

Abstract

Recently, 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 locating 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 a hypothesis for more in-depth discussion and research. Therefore, we introduce KLoB, a benchmark examining three essential properties that a reliable knowledge locating method should satisfy. KLoB can serve as a benchmark for evaluating existing locating methods in language models, and can contributes a method to reassessing the validity of locality hypothesis of factual knowledge. KLoB is publicly available at an anonymous GitHub: \url{https://github.com/anon6662/KLoB}.

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

TL;DR

The paper tackles the lack of rigorous benchmarks for evaluating knowledge locating methods in language models and the ongoing debate about the locality of factual knowledge within model parameters. It introduces KLoB, a benchmark that enforces three properties—Consistency, Relevance, and Unbiasedness—to assess locating results, and provides three subtasks (KLoB-c, KLoB-r, KLoB-u) anchored in Wikidata and MQUAKE data. KLoB also defines relative evaluation metrics, RSim and RSD, to enable fair cross-method comparisons across varying localization scopes. By offering data construction protocols and evaluation tools, the work supplies a concrete platform to benchmark locating methods and to reassess the locality hypothesis, with a public GitHub release for reproducibility and extension.

Abstract

Recently, 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 locating 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 a hypothesis for more in-depth discussion and research. Therefore, we introduce KLoB, a benchmark examining three essential properties that a reliable knowledge locating method should satisfy. KLoB can serve as a benchmark for evaluating existing locating methods in language models, and can contributes a method to reassessing the validity of locality hypothesis of factual knowledge. KLoB is publicly available at an anonymous GitHub: \url{https://github.com/anon6662/KLoB}.
Paper Structure (15 sections, 3 figures, 2 tables)

This paper contains 15 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Preliminary
  3. Knowledge Editing
  4. Locate-Then-Edit Method
  5. KLoB: Knowledge Locating Benchmark
  6. Design Philosophy
  7. Consistency
  8. Relevance
  9. Unbiasedness
  10. Data Format
  11. Data Construction
  12. Evaluation Metrics
  13. Conlusion
  14. Detailed Description for calculating RSim
  15. Templates of KLoB-c

Figures (3)

  • Figure 1: Illustration of Locate-Then-Edit method.
  • Figure 2: Examples of three subtasks in KLoB. KLoB-c: Each example comprises multiple expressions of the same factual knowledge; KLoB-r: Each example includes a sentence with specific factual knowledge and another with a related knowledge chain; KLoB-u: Each example features a sentence composed of random words.
  • Figure 3: $RSim$: evaluation metrics for KLoB-c and KLob-r. $RSim$ is given by $RSim = max(\frac{p}{q}, \ 0)$.