Does your LLM truly unlearn? An embarrassingly simple approach to recover unlearned knowledge

Zhiwei Zhang, Fali Wang, Xiaomin Li, Zongyu Wu, Xianfeng Tang, Hui Liu, Qi He, Wenpeng Yin, Suhang Wang

Introduction and Problem Statement

As Large Language Models (LLMs) scale, they inevitably ingest problematic data, including personally identifiable information (PII), copyrighted materials, and harmful instructions. Retraining these models from scratch to remove such data is computationally prohibitive. Consequently, researchers have turned to Machine Unlearning (MU)—a process designed to surgically remove specific knowledge from a pre-trained model while preserving its general utility.

However, this paper identifies a "catastrophic failure" in current unlearning paradigms. While existing methods appear successful in full-precision evaluations (FP32/BF16), the authors reveal that the "forgotten" knowledge often remains hidden within the model's weights. This vulnerability is exposed through a simple, industry-standard process: quantization.

Figure 1: The lifecycle of machine unlearning and the subsequent recovery of sensitive data through quantization.

The Quantization Recovery Attack

Quantization is typically used to compress models for deployment on consumer hardware, such as converting 32-bit floating-point weights to 4-bit integers. The authors frame this as a "recovery attack."

Existing unlearning methods often perform the bare minimum weight adjustments to satisfy forgetting criteria. Because these methods use small learning rates to avoid damaging the model's general intelligence, the resulting weight changes are minute. When the model is quantized into discrete "bins," these tiny changes are frequently rounded back to the original values of the pre-unlearned model, effectively "undoing" the deletion. The study finds that while 8-bit quantization behaves similarly to full precision, 4-bit quantization—the standard for local LLM use—consistently triggers the recovery of forgotten data.

Experimental Evidence of Failure

The researchers evaluated several state-of-the-art unlearning methods, including Gradient Ascent (GA) and Negative Preference Optimization (NPO), across benchmarks like MUSE (Machine Unlearning Six-Way Evaluation). The results were startling: across multiple methods, unlearned models retained an average of only 21% of forgotten knowledge in full precision, but this retention spiked to 83% after 4-bit quantization.

Figure 3: Performance metrics across NEWS and BOOKS datasets, demonstrating how 4-bit quantization (highlighted in red) triggers a massive spike in knowledge retention (M1/M2) compared to full precision.

As shown in the data, methods like GA_KLR on the BOOKS dataset show a jump from 13.0% retention (M1) in full precision to 75.6% in 4-bit quantization. This suggests that the knowledge was never truly deleted, but merely masked by subtle weight shifts that quantization ignores.

The Solution: SURE

To address this, the authors propose SURE (Saliency-Based Unlearning with a Large Learning Rate). The core idea is to move the "knowledge weights" far enough that they cross the quantization threshold ($\Delta$), making the unlearning permanent even after compression.

  1. Saliency Mapping: The framework identifies specific modules (layers or heads) most responsible for the targeted "forget" data by calculating gradients.
  2. Targeted Aggression: A significantly larger learning rate is applied only to these salient modules.
  3. Weight Freezing: By freezing non-essential weights, the model can push the critical weights far enough to resist quantization recovery without destroying the model's overall performance on general benchmarks like MMLU.

Conclusion

This research highlights a fundamental tension between model utility and unlearning robustness. Current safety tuning is inadequate for real-world deployment scenarios where models are compressed. The authors advocate for a new standard in unlearning benchmarks: models should only be certified as "unlearned" if they pass evaluations across multiple precision levels (FP16, INT8, and INT4). While the SURE framework provides a robust first step, the discovery of the quantization recovery attack suggests that truly removing knowledge from the neural weights of an LLM is a much more difficult task than previously assumed.

Summary updated

Abstract

Large language models (LLMs) have shown remarkable proficiency in generating text, benefiting from extensive training on vast textual corpora. However, LLMs may also acquire unwanted behaviors from the diverse and sensitive nature of their training data, which can include copyrighted and private content. Machine unlearning has been introduced as a viable solution to remove the influence of such problematic content without the need for costly and time-consuming retraining. This process aims to erase specific knowledge from LLMs while preserving as much model utility as possible. Despite the effectiveness of current unlearning methods, little attention has been given to whether existing unlearning methods for LLMs truly achieve forgetting or merely hide the knowledge, which current unlearning benchmarks fail to detect. This paper reveals that applying quantization to models that have undergone unlearning can restore the "forgotten" information. To thoroughly evaluate this phenomenon, we conduct comprehensive experiments using various quantization techniques across multiple precision levels. We find that for unlearning methods with utility constraints, the unlearned model retains an average of 21\% of the intended forgotten knowledge in full precision, which significantly increases to 83\% after 4-bit quantization. Based on our empirical findings, we provide a theoretical explanation for the observed phenomenon and propose a quantization-robust unlearning strategy to mitigate this intricate issue...
BibTeX
@inproceedings{zhang2025does,
  author = {Zhiwei Zhang and Fali Wang and Xiaomin Li and Zongyu Wu and Xianfeng Tang and Hui Liu and Qi He and Wenpeng Yin and Suhang Wang},
  title = {{Does your LLM truly unlearn? An embarrassingly simple approach to recover unlearned knowledge}},
  booktitle = {International Conference on Learning Representations},
  year = {2025},
  url = {https://openreview.net/forum?id=lHSeDYamnz},
}
Figure