Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

QUAIL: Quantization Aware Unlearning for Mitigating Misinformation in LLMs

Himanshu Mishra, Kanwal Mehreen

TL;DR

This work identifies a critical vulnerability where post-training low-bit quantization can reverse machine unlearning in LLMs, effectively restoring forgotten knowledge. It reveals that standard unlearning updates are too small to cross quantization boundaries, causing 'bucket collapse' under aggressive quantization. To counter this, the authors propose QUAIL, which augments unlearning with a logit-space hinge loss that enforces a margin ($Δ_q/2$) to ensure weight updates survive quantization, balancing forgetting with utility. The approach is validated on NEWS, BOOKS, and Twitter misinformation tasks, showing substantial improvements in forgetting persistence under 4-bit quantization while preserving retain-set performance. The results highlight quantization robustness as a essential deployment-aware dimension for unlearning methods and point to future directions like layer-wise margins and integration with advanced quantization techniques.

Abstract

Machine unlearning aims to remove specific knowledge (e.g., copyrighted or private data) from a trained model without full retraining. In practice, models are often quantized (e.g., 4-bit) for deployment, but we find that quantization can catastrophically restore forgotten information [1]. In this paper, we (1) analyze why low-bit quantization undermines unlearning, and (2) propose a quantization-aware unlearning method to mitigate this. We first compute weight-change statistics and bucket overlaps in quantization to show that typical unlearning updates are too small to cross quantization thresholds. Building on this insight, we introduce a logits space hinge loss: for each forget example, we force the output logits of the unlearned model to differ from the original model by at least a margin (half the quantization step). This ensures forgotten examples remain distinguishable even after quantization. We evaluate on language and classification tasks (including a Twitter misinformation dataset) and show our method preserves forgetting under 4-bit quantization, whereas existing methods almost entirely recover the forgotten knowledge.

QUAIL: Quantization Aware Unlearning for Mitigating Misinformation in LLMs

TL;DR

This work identifies a critical vulnerability where post-training low-bit quantization can reverse machine unlearning in LLMs, effectively restoring forgotten knowledge. It reveals that standard unlearning updates are too small to cross quantization boundaries, causing 'bucket collapse' under aggressive quantization. To counter this, the authors propose QUAIL, which augments unlearning with a logit-space hinge loss that enforces a margin () to ensure weight updates survive quantization, balancing forgetting with utility. The approach is validated on NEWS, BOOKS, and Twitter misinformation tasks, showing substantial improvements in forgetting persistence under 4-bit quantization while preserving retain-set performance. The results highlight quantization robustness as a essential deployment-aware dimension for unlearning methods and point to future directions like layer-wise margins and integration with advanced quantization techniques.

Abstract

Machine unlearning aims to remove specific knowledge (e.g., copyrighted or private data) from a trained model without full retraining. In practice, models are often quantized (e.g., 4-bit) for deployment, but we find that quantization can catastrophically restore forgotten information [1]. In this paper, we (1) analyze why low-bit quantization undermines unlearning, and (2) propose a quantization-aware unlearning method to mitigate this. We first compute weight-change statistics and bucket overlaps in quantization to show that typical unlearning updates are too small to cross quantization thresholds. Building on this insight, we introduce a logits space hinge loss: for each forget example, we force the output logits of the unlearned model to differ from the original model by at least a margin (half the quantization step). This ensures forgotten examples remain distinguishable even after quantization. We evaluate on language and classification tasks (including a Twitter misinformation dataset) and show our method preserves forgetting under 4-bit quantization, whereas existing methods almost entirely recover the forgotten knowledge.
Paper Structure (44 sections, 21 equations, 5 figures, 3 tables, 1 algorithm)

This paper contains 44 sections, 21 equations, 5 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Machine Unlearning
  4. Unlearning + Quantization
  5. Preliminaries
  6. Theoretical Explanation of Failure of Unlearning via Quantization
  7. The Quantization Bucket Collapse Problem
  8. Weight Change Bounds Under Standard Unlearning
  9. Quantization-Induced Knowledge Recovery
  10. Logit-Space Analysis
  11. Empirical Analysis of Failure of Unlearning via Quantization
  12. Experimental Setup
  13. Weight Change Statistics
  14. Quantized Bucket Overlap
  15. Layerwise Overlap Analysis
  16. ...and 29 more sections

Figures (5)

  • Figure 1: Layerwise overlap across 16 bit, 8 bit, and 4 bit precision.
  • Figure 2: Layerwise comparison between $f_{\mathrm{GA\text{-}GDR}}$ and $Q(f_{\mathrm{GA\text{-}GDR}})$.
  • Figure 3: Histogram of per tensor overlap at 16 bit precision.
  • Figure 4: Histogram of per tensor overlap at 8 bit precision.
  • Figure 5: Histogram of per tensor overlap at 4 bit precision.