# Do LLMs Really Forget? Hidden-State Leakage in Model Unlearning and How to Fix it

arXiv: https://arxiv.org/abs/2609.36612 · HTML: https://mlsanity.com/paper/2609.36612/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Machine unlearning, one of 12 directions in LLM safety and security.
- Builds on 48 of its 53 references that are in the map.
- First posted on arXiv 29 Sep 2026.

## Placed in

- LLM safety and security / [Machine unlearning](https://mlsanity.com/safety/machine-unlearning/)

## Builds on (most cited first, 30 of 48)

- [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://mlsanity.com/paper/1810.04805/) (2018)
- [Llama 2: Open Foundation and Fine-Tuned Chat Models](https://mlsanity.com/paper/2307.09288/) (2023)
- [Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks](https://mlsanity.com/paper/1908.10084/) (2019)
- [Towards Deep Learning Models Resistant to Adversarial Attacks](https://mlsanity.com/paper/1706.06083/) (2017)
- [Universal and Transferable Adversarial Attacks on Aligned Language Models](https://mlsanity.com/paper/2307.15043/) (2023)
- [Membership Inference Attacks against Machine Learning Models](https://mlsanity.com/paper/1610.05820/) (2016)
- [SimCSE: Simple Contrastive Learning of Sentence Embeddings](https://mlsanity.com/paper/2104.08821/) (2021)
- [Theoretically Principled Trade-off between Robustness and Accuracy](https://mlsanity.com/paper/1901.08573/) (2019)
- [Understanding intermediate layers using linear classifier probes](https://mlsanity.com/paper/1610.01644/) (2016)
- [The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning](https://mlsanity.com/paper/2403.03218/) (2024)
- [Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning](https://mlsanity.com/paper/2404.05868/) (2024)
- [Scalable Extraction of Training Data from (Production) Language Models](https://mlsanity.com/paper/2311.17035/) (2023)
- [Detecting Pretraining Data from Large Language Models](https://mlsanity.com/paper/2310.16789/) (2023)
- [Linguistic Knowledge and Transferability of Contextual Representations](https://mlsanity.com/paper/1903.08855/) (2019)
- [Improving Text Embeddings with Large Language Models](https://mlsanity.com/paper/2401.00368/) (2023)
- [TrustLLM: Trustworthiness in Large Language Models](https://mlsanity.com/paper/2401.05561/) (2024)
- [MUSE: Machine Unlearning Six-Way Evaluation for Language Models](https://mlsanity.com/paper/2407.06460/) (2024)
- [Fine-grained Analysis of Sentence Embeddings Using Auxiliary Prediction Tasks](https://mlsanity.com/paper/1608.04207/) (2016)
- [Gender bias and stereotypes in Large Language Models](https://mlsanity.com/paper/2308.14921/) (2023)
- [Latent Adversarial Training Improves Robustness to Persistent Harmful Behaviors in LLMs](https://mlsanity.com/paper/2407.15549/) (2024)
- [Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning](https://mlsanity.com/paper/2410.07163/) (2024)
- [Eight Methods to Evaluate Robust Unlearning in LLMs](https://mlsanity.com/paper/2402.16835/) (2024)
- [Can Sensitive Information Be Deleted From LLMs? Objectives for Defending Against Extraction Attacks](https://mlsanity.com/paper/2309.17410/) (2023)
- [Continual Learning and Private Unlearning](https://mlsanity.com/paper/2203.12817/) (2022)
- [An Adversarial Perspective on Machine Unlearning for AI Safety](https://mlsanity.com/paper/2409.18025/) (2024)
- [Copyright Violations and Large Language Models](https://mlsanity.com/paper/2310.13771/) (2023)
- [Min-K%++: Improved Baseline for Detecting Pre-Training Data from Large Language Models](https://mlsanity.com/paper/2404.02936/) (2024)
- [Identifying and Mitigating the Security Risks of Generative AI](https://mlsanity.com/paper/2308.14840/) (2023)
- [Scaling Sentence Embeddings with Large Language Models](https://mlsanity.com/paper/2307.16645/) (2023)
- [Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning](https://mlsanity.com/paper/2406.13356/) (2024)

## Abstract (arXiv)

Unlearning in large language models (LLMs) is typically evaluated at the output level, where a model appears to suppress sensitive or undesirable content. In this work, we show that such evaluations can create an illusion of forgetting: even when output-level leakage is eliminated, sensitive information can remain encoded in the model's hidden representations. We first provide a theoretical analysis establishing a fundamental separation between output suppression and representational erasure. Specifically, we show that the decoder can be made arbitrarily insensitive to sensitive directions, driving output-level leakage to zero, while the hidden representations retain the underlying information. To empirically validate this phenomenon, we train generative probe decoders on hidden states across transformer layers, enabling layer-wise measurement of information leakage. Across three widely used benchmarks, TOFU, MUSE, and WMDP, and state-of-the-art unlearning methods, we find that substantial sensitive information remains recoverable from hidden representations, even when standard output-level metrics indicate successful unlearning. To address this gap, we propose Probe-Adversarial Representation Suppression (PARS), an unlearning objective that adversarially minimizes the extractable information from hidden representations. PARS directly targets representational leakage and provides significantly stronger guarantees of erasure under adversarial probing and relearning attacks, outperforming all evaluated baselines. Our results highlight a fundamental limitation of existing unlearning paradigms and suggest that true forgetting in LLMs requires controlling not only model outputs, but also the information encoded in hidden representations. Codes are available at https://github.com/OptimAI-Lab/HiddenStateUnlearning.