# MeGU: Machine-Guided Unlearning with Target Feature Disentanglement Source: [MeGU: Machine-Guided Unlearning with Target Feature Disentanglement]() Collection index: [llms.txt]() Data updated: 2026-02-21T05:00:04.173Z Authors: Haoyu Wang, Zhuo Huang, Xiaolong Wang, Bo Han, Zhiwei Lin, Tongliang Liu Published: 2026-02-19T05:20:31.000Z ## Abstract The growing concern over training data privacy has elevated the "Right to be Forgotten" into a critical requirement, thereby raising the demand for effective Machine Unlearning. However, existing unlearning approaches commonly suffer from a fundamental trade-off: aggressively erasing the influence of target data often degrades model utility on retained data, while conservative strategies leave residual target information intact. In this work, the intrinsic representation properties learned during model pretraining are analyzed. It is demonstrated that semantic class concepts are entangled at the feature-pattern level, sharing associated features while preserving concept-specific discriminative components. This entanglement fundamentally limits the effectiveness of existing unlearning paradigms. Motivated by this insight, we propose Machine-Guided Unlearning (MeGU), a novel framework that guides unlearning through concept-aware re-alignment. Specifically, Multi-modal Large Language Models (MLLMs) are leveraged to explicitly determine re-alignment directions for target samples by assigning semantically meaningful perturbing labels. To improve efficiency, inter-class conceptual similarities estimated by the MLLM are encoded into a lightweight transition matrix. Furthermore, MeGU introduces a positive-negative feature noise pair to explicitly disentangle target concept influence. During finetuning, the negative noise suppresses target-specific feature patterns, while the positive noise reinforces remaining associated features and aligns them with perturbing concepts. This coordinated design enables selective disruption of target-specific representations while preserving shared semantic structures. As a result, MeGU enables controlled and selective forgetting, effectively mitigating both under-unlearning and over-unlearning. ## Recorded benchmark results A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage. - [Sub-class Machine Unlearning on CIFAR-20](): Sub-class Machine Unlearning; CIFAR-20; Ar: 83.1. - [Random Unlearning on CIFAR-10 (original)](): Random Unlearning; CIFAR-10 original; Ar: 98.81. - [Machine Unlearning on CIFAR-100 class baby (test)](): Machine Unlearning; CIFAR-100 class baby (test); Utility (Ar): 76.7. - [Machine Unlearning on CIFAR-100 class DINO (test)](): Machine Unlearning; CIFAR-100 class DINO (test); Utility Ar: 76.55. - [Class-wise Machine Unlearning on CIFAR-20 veh2 (test)](): Class-wise Machine Unlearning; CIFAR-20 veh2 (test); Ar Utility Score: 95.82. - [Class-wise Machine Unlearning on CIFAR-20 veg (test)](): Class-wise Machine Unlearning; CIFAR-20 veg (test); Ar (Accuracy Retention): 95.64. - [Machine Unlearning on CIFAR-100 class wolf (test)](): Machine Unlearning; CIFAR-100 class wolf (test); Accuracy Retention (Ar): 76.3. - [Sub-class Unlearning on CIFAR-20 (test)](): Sub-class Unlearning; CIFAR-20 (test); Ar: 96.05. - [Machine Unlearning on CIFAR-100 class lamp (test)](): Machine Unlearning; CIFAR-100 class lamp (test); Ar (Accuracy Retention): 76.11. - [Machine Unlearning on CIFAR-100 class sea (test)](): Machine Unlearning; CIFAR-100 class sea (test); Utility (Ar): 73.36. - [Machine Unlearning on CIFAR-100 (test)](): Machine Unlearning; CIFAR-100 (test); Retain Acc: 92.34. ## Original sources and implementation [Original paper]() [PDF]()