Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

On the Robustness of Editing Large Language Models

Xinbei Ma, Tianjie Ju, Jiyang Qiu, Zhuosheng Zhang, Hai Zhao, Lifeng Liu, Yulong Wang

TL;DR

The authors' empirical studies uncover a substantial disparity between existing editing methods and the practical application of LLMs, and shows that more popular knowledge is memorized better, easier to recall, and more challenging to edit effectively.

Abstract

Large language models (LLMs) have played a pivotal role in building communicative AI, yet they encounter the challenge of efficient updates. Model editing enables the manipulation of specific knowledge memories and the behavior of language generation without retraining. However, the robustness of model editing remains an open question. This work seeks to understand the strengths and limitations of editing methods, facilitating practical applications of communicative AI. We focus on three key research questions. RQ1: Can edited LLMs behave consistently resembling communicative AI in realistic situations? RQ2: To what extent does the rephrasing of prompts lead LLMs to deviate from the edited knowledge memory? RQ3: Which knowledge features are correlated with the performance and robustness of editing? Our empirical studies uncover a substantial disparity between existing editing methods and the practical application of LLMs. On rephrased prompts that are flexible but common in realistic applications, the performance of editing experiences a significant decline. Further analysis shows that more popular knowledge is memorized better, easier to recall, and more challenging to edit effectively. Code is publicly available at https://github.com/xbmxb/edit_analysis .

On the Robustness of Editing Large Language Models

TL;DR

The authors' empirical studies uncover a substantial disparity between existing editing methods and the practical application of LLMs, and shows that more popular knowledge is memorized better, easier to recall, and more challenging to edit effectively.

Abstract

Large language models (LLMs) have played a pivotal role in building communicative AI, yet they encounter the challenge of efficient updates. Model editing enables the manipulation of specific knowledge memories and the behavior of language generation without retraining. However, the robustness of model editing remains an open question. This work seeks to understand the strengths and limitations of editing methods, facilitating practical applications of communicative AI. We focus on three key research questions. RQ1: Can edited LLMs behave consistently resembling communicative AI in realistic situations? RQ2: To what extent does the rephrasing of prompts lead LLMs to deviate from the edited knowledge memory? RQ3: Which knowledge features are correlated with the performance and robustness of editing? Our empirical studies uncover a substantial disparity between existing editing methods and the practical application of LLMs. On rephrased prompts that are flexible but common in realistic applications, the performance of editing experiences a significant decline. Further analysis shows that more popular knowledge is memorized better, easier to recall, and more challenging to edit effectively. Code is publicly available at https://github.com/xbmxb/edit_analysis .
Paper Structure (40 sections, 1 equation, 9 figures, 7 tables)

This paper contains 40 sections, 1 equation, 9 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Model Editing
  4. Reflections on Model Editing
  5. Communicative AI
  6. Task Formulation
  7. RQ1: Edited LLM as communicative AI
  8. Method
  9. Experiments for RQ1
  10. Analysis for RQ1
  11. RQ2: "Attack" for Editing
  12. Method
  13. Experiments for RQ2
  14. Datasets
  15. Baselines and Implementation
  16. ...and 25 more sections

Figures (9)

  • Figure 1: Overview of our work. The upper part illustrates the editing success on target knowledge (Section \ref{['31']}). The lower part shows our studies on the edited model in realistic use. The left part shows the risks of edited LLMs as communicative AI (Section \ref{['32']}) and the right part shows our "attack" for editing (Section \ref{['33']}).
  • Figure 2: Edited communicative AI. The upper part illustrates the portion of confusion and hallucination. The bottom shows a case that appears knowledge reversion.
  • Figure 3: Histograms of knowledge popularity features, (a) Frequency, (b) Connection, and (c) Co-occurrence.
  • Figure 4: Editing performance on different levels of (a) Frequency, (b) Connection, and (c) Co-occurrence.
  • Figure 5: A case of human evaluation.
  • ...and 4 more figures