Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LLM Unlearning via Neural Activation Redirection

William F. Shen, Xinchi Qiu, Meghdad Kurmanji, Alex Iacob, Lorenzo Sani, Yihong Chen, Nicola Cancedda, Nicholas D. Lane

TL;DR

This work tackles the problem of removing specific information from large language models without sacrificing general performance and while maintaining controllability over outputs. It introduces LUNAR, an activation-redirection method that selectively redirects forget-set activations to regions encoding ignorance, using a single down-projection matrix with a closed-form solution for training. Across multiple base models and datasets, LUNAR achieves state-of-the-art unlearning efficacy and superior controllability, and remains robust under adversarial attacks and sequential unlearning tasks. The approach blends mechanistic interpretability with practical unlearning, offering significant efficiency gains and compatibility with PEFT techniques like LoRA for scalable deployment.

Abstract

The ability to selectively remove knowledge from LLMs is highly desirable. However, existing methods often struggle with balancing unlearning efficacy and retain model utility, and lack controllability at inference time to emulate base model behavior as if it had never seen the unlearned data. In this paper, we propose LUNAR, a novel unlearning method grounded in the Linear Representation Hypothesis and operates by redirecting the representations of unlearned data to activation regions that expresses its inability to answer. We show that contrastive features are not a prerequisite for effective activation redirection, and LUNAR achieves state-of-the-art unlearning performance and superior controllability. Specifically, LUNAR achieves between 2.9x and 11.7x improvement in the combined unlearning efficacy and model utility score (Deviation Score) across various base models and generates coherent, contextually appropriate responses post-unlearning. Moreover, LUNAR effectively reduces parameter updates to a single down-projection matrix, a novel design that significantly enhances efficiency by 20x and robustness. Finally, we demonstrate that LUNAR is robust to white-box adversarial attacks and versatile in real-world scenarios, including handling sequential unlearning requests.

LLM Unlearning via Neural Activation Redirection

TL;DR

This work tackles the problem of removing specific information from large language models without sacrificing general performance and while maintaining controllability over outputs. It introduces LUNAR, an activation-redirection method that selectively redirects forget-set activations to regions encoding ignorance, using a single down-projection matrix with a closed-form solution for training. Across multiple base models and datasets, LUNAR achieves state-of-the-art unlearning efficacy and superior controllability, and remains robust under adversarial attacks and sequential unlearning tasks. The approach blends mechanistic interpretability with practical unlearning, offering significant efficiency gains and compatibility with PEFT techniques like LoRA for scalable deployment.

Abstract

The ability to selectively remove knowledge from LLMs is highly desirable. However, existing methods often struggle with balancing unlearning efficacy and retain model utility, and lack controllability at inference time to emulate base model behavior as if it had never seen the unlearned data. In this paper, we propose LUNAR, a novel unlearning method grounded in the Linear Representation Hypothesis and operates by redirecting the representations of unlearned data to activation regions that expresses its inability to answer. We show that contrastive features are not a prerequisite for effective activation redirection, and LUNAR achieves state-of-the-art unlearning performance and superior controllability. Specifically, LUNAR achieves between 2.9x and 11.7x improvement in the combined unlearning efficacy and model utility score (Deviation Score) across various base models and generates coherent, contextually appropriate responses post-unlearning. Moreover, LUNAR effectively reduces parameter updates to a single down-projection matrix, a novel design that significantly enhances efficiency by 20x and robustness. Finally, we demonstrate that LUNAR is robust to white-box adversarial attacks and versatile in real-world scenarios, including handling sequential unlearning requests.

Paper Structure

This paper contains 33 sections, 3 theorems, 30 equations, 9 figures, 16 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. LUNAR
  4. Unlearning via Neural Activation Redirection
  5. Layer Selection
  6. Memory and Computational Costs
  7. Analytical Solution and Convergence Study
  8. Experiment Setup
  9. Experimental Setup
  10. Results
  11. Unlearning Performance
  12. Unlearning Pre-trained Data from Base Models
  13. Unlearning Sequentially
  14. Robustness Study
  15. Related Works
  16. ...and 18 more sections

Key Result

Lemma 1

Let $[H_f, H_r] \in \mathbb{R}^{m \times n}$ (with $m \geq n$). The Gram matrix $[H_f, H_r]^\top [H_f, H_r]$ is invertible if and only if the columns of $[H_f, H_r]$ are linearly independent.

Figures (9)

  • Figure 2: PCA visualization of activation space post LUNAR unlearning: (a) unlearn edge AB from the PISTOL dataset; (b) unlearn the first author from the TOFU dataset; (c) unlearn factual dataset from base model with reference dataset be the harmful dataset; (d) unlearn factual dataset from base model with reference dataset be the unverifiable dataset. Base model and PISTOL/TOFU SFT models are Llama2-based.
  • Figure 3: Analysis of the most effective layer for activation redirection.
  • Figure : (a)
  • Figure : (a)
  • Figure : (b)
  • ...and 4 more figures

Theorems & Definitions (8)

  • Lemma 1
  • proof
  • Lemma 2: Convexity
  • proof
  • Lemma 3: Lipschitz Continuity of Gradient
  • Definition 1
  • proof
  • Remark 1