Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

How Data Inter-connectivity Shapes LLMs Unlearning: A Structural Unlearning Perspective

Xinchi Qiu, William F. Shen, Yihong Chen, Meghdad Kurmanji, Nicola Cancedda, Pontus Stenetorp, Nicholas D. Lane

TL;DR

This work tackles the gap in LLM unlearning by introducing PISTOL, a pipeline that constructs structured, contract-based datasets to study forgetting in the presence of data inter-connectivity. It demonstrates that increased inter-connectivity and higher knowledge-graph density make unlearning harder, and that domain-skew in forget data can disproportionately affect retention in the same domain. The paper compares gradient-based and preference-optimization unlearning methods, finding PO approaches more robust to structural effects, and shows that inter-connectivity extends to pre-training data as well. Overall, PISTOL provides a controllable benchmark that reveals structural factors shaping unlearning effectiveness and offers a clearer evaluation framework free from pre-training confounds, with practical implications for policy-driven data erasure and safer LLM deployment.

Abstract

While unlearning knowledge from large language models (LLMs) is receiving increasing attention, one important aspect remains unexplored. Existing approaches and benchmarks assume data points to-be-forgotten are independent, ignoring their inter-connectivity - a fundamental characteristic of real-world data structures. In this paper, we propose PISTOL, a method for compiling structural datasets. PISTOL leverages the inherently structured nature of contractual relationships, offering several key benefits. First, it enables insights into the impact of structural data on unlearning effectiveness. Second, it provides precise and concise ground truths for clearer evaluation. Third, its attribute generation does not require input from pre-trained LLMs, mitigating confounding risks. Leveraging datasets synthesized using PISTOL, we demonstrate how data inter-connectivity impacts LLM unlearning. Specifically, (a) in both the pre-trained and fine-tuned models, unlearning difficulty increases as data inter-connectivity grows, (b) there is a positive correlation between the density of the knowledge graph and unlearning difficulty, and (c) when the to-be-forgotten data is skewed towards one domain, balancing retaining performance across all domains is challenging.

How Data Inter-connectivity Shapes LLMs Unlearning: A Structural Unlearning Perspective

TL;DR

This work tackles the gap in LLM unlearning by introducing PISTOL, a pipeline that constructs structured, contract-based datasets to study forgetting in the presence of data inter-connectivity. It demonstrates that increased inter-connectivity and higher knowledge-graph density make unlearning harder, and that domain-skew in forget data can disproportionately affect retention in the same domain. The paper compares gradient-based and preference-optimization unlearning methods, finding PO approaches more robust to structural effects, and shows that inter-connectivity extends to pre-training data as well. Overall, PISTOL provides a controllable benchmark that reveals structural factors shaping unlearning effectiveness and offers a clearer evaluation framework free from pre-training confounds, with practical implications for policy-driven data erasure and safer LLM deployment.

Abstract

While unlearning knowledge from large language models (LLMs) is receiving increasing attention, one important aspect remains unexplored. Existing approaches and benchmarks assume data points to-be-forgotten are independent, ignoring their inter-connectivity - a fundamental characteristic of real-world data structures. In this paper, we propose PISTOL, a method for compiling structural datasets. PISTOL leverages the inherently structured nature of contractual relationships, offering several key benefits. First, it enables insights into the impact of structural data on unlearning effectiveness. Second, it provides precise and concise ground truths for clearer evaluation. Third, its attribute generation does not require input from pre-trained LLMs, mitigating confounding risks. Leveraging datasets synthesized using PISTOL, we demonstrate how data inter-connectivity impacts LLM unlearning. Specifically, (a) in both the pre-trained and fine-tuned models, unlearning difficulty increases as data inter-connectivity grows, (b) there is a positive correlation between the density of the knowledge graph and unlearning difficulty, and (c) when the to-be-forgotten data is skewed towards one domain, balancing retaining performance across all domains is challenging.

Paper Structure

This paper contains 36 sections, 6 equations, 5 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Limitations of Existing Datasets
  3. PISTOL Dataset
  4. Dataset Compilation Pipeline
  5. Dataset
  6. Dataset 1.
  7. Dataset 2.
  8. Evaluation Setup
  9. Metrics.
  10. Unlearning baselines
  11. Base models.
  12. Evaluation Methods
  13. Impact of data inter-connectivity.
  14. Impact of data domain.
  15. Results
  16. ...and 21 more sections

Figures (5)

  • Figure 1: Illustration of structural unlearning (i.e., unlearning inter-connected data points within a structured dataset) versus independent unlearning (i.e., unlearning isolated data points). As shown, when an entity revokes consent for its data to be used or exercises its 'right to be forgotten' (i.e., unlearning data points related to this entity), the degree of inter-connectivity between the unlearning entity and other entities will influence unlearning performance.
  • Figure 2: Illustration of the Dataset Compilation Pipeline.
  • Figure : (a)
  • Figure : (a)
  • Figure : (b)