Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Unlearning Mirage: A Dynamic Framework for Evaluating LLM Unlearning

Raj Sanjay Shah, Jing Huang, Keerthiram Murugesan, Nathalie Baracaldo, Diyi Yang

TL;DR

A dynamic framework that stress tests unlearning robustness using complex structured queries, enabling practical and scalable evaluation of unlearning methods without the need for manual construction of forget test sets, enabling easier adoption for real-world applications.

Abstract

Unlearning in Large Language Models (LLMs) aims to enhance safety, mitigate biases, and comply with legal mandates, such as the right to be forgotten. However, existing unlearning methods are brittle: minor query modifications, such as multi-hop reasoning and entity aliasing, can recover supposedly forgotten information. As a result, current evaluation metrics often create an illusion of effectiveness, failing to detect these vulnerabilities due to reliance on static, unstructured benchmarks. We propose a dynamic framework that stress tests unlearning robustness using complex structured queries. Our approach first elicits knowledge from the target model (pre-unlearning) and constructs targeted probes, ranging from simple queries to multi-hop chains, allowing precise control over query difficulty. Our experiments show that the framework (1) shows comparable coverage to existing benchmarks by automatically generating semantically equivalent Q&A probes, (2) aligns with prior evaluations, and (3) uncovers new unlearning failures missed by other benchmarks, particularly in multi-hop settings. Furthermore, activation analyses show that single-hop queries typically follow dominant computation pathways, which are more likely to be disrupted by unlearning methods. In contrast, multi-hop queries tend to use alternative pathways that often remain intact, explaining the brittleness of unlearning techniques in multi-hop settings. Our framework enables practical and scalable evaluation of unlearning methods without the need for manual construction of forget test sets, enabling easier adoption for real-world applications. We release the pip package and the code at https://sites.google.com/view/unlearningmirage/home.

The Unlearning Mirage: A Dynamic Framework for Evaluating LLM Unlearning

TL;DR

A dynamic framework that stress tests unlearning robustness using complex structured queries, enabling practical and scalable evaluation of unlearning methods without the need for manual construction of forget test sets, enabling easier adoption for real-world applications.

Abstract

Unlearning in Large Language Models (LLMs) aims to enhance safety, mitigate biases, and comply with legal mandates, such as the right to be forgotten. However, existing unlearning methods are brittle: minor query modifications, such as multi-hop reasoning and entity aliasing, can recover supposedly forgotten information. As a result, current evaluation metrics often create an illusion of effectiveness, failing to detect these vulnerabilities due to reliance on static, unstructured benchmarks. We propose a dynamic framework that stress tests unlearning robustness using complex structured queries. Our approach first elicits knowledge from the target model (pre-unlearning) and constructs targeted probes, ranging from simple queries to multi-hop chains, allowing precise control over query difficulty. Our experiments show that the framework (1) shows comparable coverage to existing benchmarks by automatically generating semantically equivalent Q&A probes, (2) aligns with prior evaluations, and (3) uncovers new unlearning failures missed by other benchmarks, particularly in multi-hop settings. Furthermore, activation analyses show that single-hop queries typically follow dominant computation pathways, which are more likely to be disrupted by unlearning methods. In contrast, multi-hop queries tend to use alternative pathways that often remain intact, explaining the brittleness of unlearning techniques in multi-hop settings. Our framework enables practical and scalable evaluation of unlearning methods without the need for manual construction of forget test sets, enabling easier adoption for real-world applications. We release the pip package and the code at https://sites.google.com/view/unlearningmirage/home.
Paper Structure (42 sections, 7 equations, 4 figures, 11 tables)

This paper contains 42 sections, 7 equations, 4 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related works
  3. Evaluating unlearning in LLMs
  4. Dynamic graph-based evaluation LLMs
  5. Preliminaries
  6. Unlearning Objectives
  7. Unlearning Methods
  8. A Dynamic Evaluation Framework for Unlearning
  9. Knowledge Graph Construction
  10. Graph Expansion
  11. BFS with Decay for Expansion Control
  12. Alias Resolution via LLM Calls
  13. Additional considerations
  14. Structured Probe Generation
  15. Evaluation Protocol
  16. ...and 27 more sections

Figures (4)

  • Figure 1: Overview of our framework: Our evaluation framework constructs a knowledge graph from pre-unlearning model outputs, enabling the automatic generation of structured single-hop, multi-hop, and alias-based queries. After applying unlearning, we probe the model to assess residual knowledge. The framework is dynamic, instantiable for any entity, and structured, providing fine-grained control over query complexity.
  • Figure 2: Limitations of existing "single-hop" evaluation metrics in assessing LLM unlearning robustness. Single-hop queries might suggest successful forgetting, but minor variations, such as multi-hop reasoning or entity aliasing, can still recover the supposedly forgotten information.
  • Figure 3: For RWKU, we compare default 2-hop queries with two variants: (+ Decomposition) prompting the model to solve the query step-by-step, and (+ Aliasing) substituting intermediate entities with known aliases.
  • Figure 4: Coverage of existing benchmarks at different graph expansion rates.