Unlearning Isn't Deletion: Investigating Reversibility of Machine Unlearning in LLMs

Unlearning in large language models (LLMs) is intended to remove the influence of specific data, yet current evaluations rely heavily on token-level metrics such as accuracy and perplexity. We show that these metrics can be misleading: models often appear to forget, but their original behavior can be rapidly restored with minimal fine-tuning, revealing that unlearning may obscure information rather than erase it. To diagnose this phenomenon, we introduce a representation-level evaluation framework using PCA-based similarity and shift, centered kernel alignment, and Fisher information. Applying this toolkit across six unlearning methods, three domains (text, code, math), and two open-source LLMs, we uncover a critical distinction between reversible and irreversible forgetting. In reversible cases, models suffer token-level collapse yet retain latent features; in irreversible cases, deeper representational damage occurs. We further provide a theoretical account linking shallow weight perturbations near output layers to misleading unlearning signals, and show that reversibility is modulated by task type and hyperparameters. Our findings reveal a fundamental gap in current evaluation practices and establish a new diagnostic foundation for trustworthy unlearning in LLMs. We provide a unified toolkit for analyzing LLM representation changes under unlearning and relearning: https://github.com/XiaoyuXU1/Representational_Analysis_Tools.git.

OBLIVIATE: Robust and Practical Machine Unlearning for Large Language Models

Large language models (LLMs) trained over extensive corpora risk memorizing sensitive, copyrighted, or toxic content. To address this, we propose OBLIVIATE, a robust unlearning framework that removes targeted data while preserving model utility. The framework follows a structured process: extracting target tokens, building retain sets, and fine-tuning with a tailored loss function comprising three components -- masking, distillation, and world fact. Using low-rank adapters (LoRA), it ensures efficiency without compromising unlearning quality. We conduct experiments on multiple datasets, including the Harry Potter series, WMDP, and TOFU, using a comprehensive suite of metrics: forget quality (new document-level memorization score), model utility, and fluency. Results demonstrate its effectiveness in resisting membership inference attacks, minimizing the impact on retained data, and maintaining robustness across diverse scenarios.

Weak-To-Strong Backdoor Attacks for LLMs with Contrastive Knowledge Distillation

Despite being widely applied due to their exceptional capabilities, Large Language Models (LLMs) have been proven to be vulnerable to backdoor attacks. These attacks introduce targeted vulnerabilities into LLMs by poisoning training samples and full-parameter fine-tuning. However, this kind of backdoor attack is limited since they require significant computational resources, especially as the size of LLMs increases. Besides, parameter-efficient fine-tuning (PEFT) offers an alternative but the restricted parameter updating may impede the alignment of triggers with target labels. In this study, we first verify that backdoor attacks with PEFT may encounter challenges in achieving feasible performance. To address these issues and improve the effectiveness of backdoor attacks with PEFT, we propose a novel backdoor attack algorithm from weak to strong based on contrastive knowledge distillation (W2SAttack). Specifically, we poison small-scale language models through full-parameter fine-tuning to serve as the teacher model. The teacher model then covertly transfers the backdoor to the large-scale student model through contrastive knowledge distillation, which employs PEFT. Theoretical analysis reveals that W2SAttack has the potential to augment the effectiveness of backdoor attacks. We demonstrate the superior performance of W2SAttack on classification tasks across four language models, four backdoor attack algorithms, and two different architectures of teacher models. Experimental results indicate success rates close to 100% for backdoor attacks targeting PEFT.