Paper deep dive
Ready2Unlearn: A Learning-Time Approach for Preparing Models with Future Unlearning Readiness
Hanyu Duan, Yi Yang, Ahmed Abbasi, Kar Yan Tam
Models: GPT-2, GPT-2 Medium, LLaMA-3.2-1B, LLaMA-3.2-3B
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 3/12/2026, 5:22:05 PM
Summary
Ready2Unlearn is a proactive, learning-time optimization approach that prepares machine learning models for future unlearning requests. By leveraging meta-learning principles (specifically MAML), it optimizes model parameters during training to be 'unlearning-ready,' allowing for more efficient and reliable gradient ascent-based unlearning. The method categorizes data into 'revocable' and 'stable' groups, ensuring that models can effectively remove specific data imprints while preserving overall utility and resisting the inadvertent recovery of erased information.
Entities (5)
Relation Signals (3)
Ready2Unlearn → supports → Gradient Ascent
confidence 95% · Ready2Unlearn is model-agnostic and compatible with any gradient ascent-based machine unlearning algorithms.
Ready2Unlearn → utilizes → MAML
confidence 95% · Ready2Unlearn adopts meta-learning principles, particularly the MAML algorithm.
Ready2Unlearn → optimizes → Machine Unlearning
confidence 90% · Ready2Unlearn is a learning-time optimization approach designed to facilitate future unlearning processes.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Machine unlearning is the process of removing the imprint left by specific data samples during the training of a machine learning model. AI developers, including those building personalized technologies, employ machine unlearning for various purposes such as privacy protection, security, and to address ethical concerns. This paper introduces Ready2Unlearn, a learning-time optimization approach designed to facilitate future unlearning processes. Unlike the majority of existing unlearning efforts that focus on designing unlearning algorithms, which are typically implemented reactively when an unlearning request is made during the model deployment phase, Ready2Unlearn shifts the focus to the training phase, adopting a "forward-looking" perspective. Building upon well-established meta-learning principles, Ready2Unlearn proactively trains machine learning models with unlearning readiness, such that they are well prepared and can handle future unlearning requests in a more efficient and principled manner. Ready2Unlearn is model-agnostic and compatible with any gradient ascent-based machine unlearning algorithms. We evaluate the method on both language and vision tasks under various unlearning settings, including class-wise unlearning and random data unlearning. Experimental results show that by incorporating such preparedness at training time, Ready2Unlearn produces an unlearning-ready model state, which offers several key advantages when future unlearning is requested. We hope this study inspires future research on proactive strategies for equipping machine learning models with built-in unlearning readiness, particularly in modern information systems that rely heavily on user data for recommendation, search, and personalized services, where privacy risks and data deletion demands are increasingly prevalent.
Tags
Links
- Source: https://arxiv.org/abs/2505.10845
- Canonical: https://arxiv.org/abs/2505.10845
Trouble viewing inline? Open PDF directly →
Full Text
89,375 characters extracted from source content.
Expand or collapse full text
Ready2Unlearn: A Learning-Time Approach for Preparing Models with Future Unlearning Readiness HANYU DUAN,Hong Kong University of Science and Technology, Hong Kong YI YANG,Hong Kong University of Science and Technology, Hong Kong AHMED ABBASI,University of Notre Dame, USA KAR YAN TAM,Hong Kong University of Science and Technology, Hong Kong Machine unlearning is the process of removing the imprint left by specic data samples during the training of a machine learning model. AI developers, including those building personalized technologies, employ machine unlearning for various purposes such as privacy protection, security, and to address ethical concerns. This paper introduces Ready2Unlearn, a learning-time optimization approach designed to facilitate future unlearning processes. Unlike the majority of existing unlearning eorts that focus on designing unlearning algorithms, which are typically implemented reactively when an unlearning request is made during the model deployment phase, Ready2Unlearn shifts the focus to the training phase, adopting a“forward-looking”perspective. Building upon well-established meta-learning principles, Ready2Unlearn proactively trains machine learning models withunlearning readiness, such that they are well prepared and can handle future unlearning requests in a more ecient and principled manner. Ready2Unlearn is model-agnostic and compatible with any gradient ascent-based machine unlearning algorithms. We evaluate the method on both language and vision tasks under various unlearning settings, including class-wise unlearning and random data unlearning. Experimental results show that by incorporating such preparedness at training time, Ready2Unlearn produces anunlearning-ready model state, which oers several key advantages when future unlearning is requested, including reduced unlearning time, improved retention of overall model capability, and enhanced resistance to the inadvertent recovery of erased information. We hope this study inspires future research on proactive strategies for equipping machine learning models with built-in unlearning readiness, particularly in modern information systems that rely heavily on user data for recommendation, search, and personalized services, where privacy risks and data deletion demands are increasingly prevalent. CCS Concepts:•Computing methodologies!Machine learning;•Security and privacy!Privacy protections; Information accountability and usage control. Additional Key Words and Phrases: Machine unlearning, meta-learning, unlearning-aware learning, data deletion, AI governance ACM Reference Format: Hanyu Duan, Yi Yang, Ahmed Abbasi, and Kar Yan Tam. 2026. Ready2Unlearn: A Learning-Time Approach for Preparing Models with Future Unlearning Readiness.ACM Trans. Inf. Syst.1, 1 (March 2026),26pages. https://doi.org/10.1145/n.n 1 Introduction Machine unlearning [7] refers to the process of removing the imprint left by specic data samples during the training of a machine learning model. AI developers employ machine unlearning for various purposes. In the context of privacy protection, it is often necessary to remove the inuence that individuals’ personal data has had on a model’s learned parameters [34,61,96]. Legal frameworks such as the European Union’s General Data Protection Regulation (GDPR) [64] Authors’ Contact Information: Hanyu Duan, Hong Kong University of Science and Technology, Hong Kong, Hong Kong, hduanac@connect.ust.hk; Yi Yang, Hong Kong University of Science and Technology, Hong Kong, Hong Kong, imyiyang@ ust.hk; Ahmed Abbasi, University of Notre Dame, Notre Dame, USA, aabbasi@nd.edu; Kar Yan Tam, Hong Kong University of Science and Technology, Hong Kong, Hong Kong, kytam@ust.hk. 2026. ACM 1558-2868/2026/3-ART https://doi.org/10.1145/n.n ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. 2Hanyu Duan, Yi Yang, Ahmed Abbasi, and Kar Yan Tam and the California Consumer Privacy Act (CCPA) 1 grant individuals the right to control their personal data, including revoking them from organizations that use the data to train models. Beyond privacy, machine unlearning has also been used to address ethical and security concerns by removing the inuence of harmful or sensitive data [69,89], such as preventing large language models (LLMs) from retaining information that could be misused for developing bioweapons or launching cyberattacks [3,45,68], and to improve model performance by eliminating the impact of low-quality or noisy training samples [7,78]. Moreover, machine unlearning is increasingly relevant in personalization settings, particularly in recommender systems (RS), where user–item interaction data may need to be removed for various reasons, such as to comply with legal requirements [9,32,49,50,90], mitigate selection bias [44], reduce the risk of sensitive information leakage [10], or refresh outdated user preference proles [16]. These diverse applications emphasize that machine unlearning is important and practically meaningful. Numerous unlearning algorithms have been introduced in recent years, employing techniques such as ensemble retraining [6,9,49,50], log-based rollback [32,90], preference optimization [59,92], gradient rectication [28,53,83], and data augmentation [8,59,63], to name a few. Despite these eorts, unlearning remains a challenging task. First, it often requires considerable time or a large number of optimization steps to achieve satisfactory forgetting, especially for large- scale models such as LLMs [20]. Additionally, balancing the trade-obetween forgetting data and preserving overall model utility is dicult [28,53,83], as the unlearning process can lead to catastrophic forgetting [92]. Moreover, some studies suggest that current unlearning methods may not be as reliable as they appear, with data that seems to be “forgotten” often being easily recovered [29,38,93]. Such long-standing challenges prompt us to ask:Is the model truly ready to forget when unlearning is initiated, and can we take steps during training to proactively prepare it with unlearning readiness against potential future unlearning requests?In this paper, we explore the possibility that equipping the model with unlearning readiness during the learning phase to benet the unlearning process that may take place later after model deployment, with improved unlearning eciency and reliability. The problem is illustrated in Figure1. We make a practical assumption that, in real-world applications, not all training data is equally likely to be subject to future unlearning requests [39]. Some data, such as user-generated content (UGC) [18], are more likely to be revoked because of privacy or other regulatory concerns, while other data, like public datasets (e.g., Wikipedia dumps), are less likely to trigger such requests. For example, in information retrieval (IR) systems, time-sensitive content, such as news articles or event-related information, is more likely to require updating or removal as it becomes outdated, whereas general world knowledge or established reference information tends to remain stable and is rarely subject to revocation [46,97]. Similarly, in personalized recommendation, interaction data from highly engaged users with rapidly evolving interests are more likely to require updates or unlearning. By contrast, users with stable, long-term preferences or those likely to churn from the service typically generate data that are less prone to revocation [9,10,16,32]. Importantly, in practice, such heterogeneity in future data modication likelihood can often be inferred using established customer analytics and user behavior modeling techniques. Motivated by this observation, we introduce a method (Ready2Unlearn) that equips the model, during training, with an awareness of potential unlearning actions that may occur later during deployment for data that are more likely to be revoked, which we termrevocable datathroughout the manuscript. The remaining data are therefore termedstable data,reecting their inherent stability, that is, their low likelihood of being subject to future revocation requests. In terms of 1 https://oag.ca.gov/privacy/ccpa ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. Ready2Unlearn: A Learning-Time Approach for Preparing Models with Future Unlearning Readiness3 Fig. 1. Comparison of learning with (top) and without (boom) unlearning preparation. which data should be classied as revocable or stable, we consider this to be largely a managerial decision, guided by organizational priorities and regulatory considerations. For example, in many enterprise settings subject to GDPR, customer data are routinely classied by their purpose (e.g., marketing, onboarding) and source (e.g., publicly accessible information, user-provided data). 2 We argue that such operational data-governance practices naturally provide insight into which categories are more likely to be revoked (e.g., user-provided data) and which are generally stable (e.g., publicly accessible information), and therefore, in this work, we do not prescribe a one-size- ts-all algorithmic approach for this categorization, as it is best determined by practitioners based on domain-specic knowledge and institutional policies. Our expectation is that, once a model has undergone learning with Ready2Unlearn and is deployed in the environment, and an unlearning request arises for the revocable data, it can 1) unlearn more eciently with fewer unlearning steps, 2) better preserve performance on the unaected stable data, and 3) exhibit greater post-unlearning robustness against the inadvertent recovery ofpreviously erased information, compared to a model trained without unlearning readiness (the naïve model). Two points to note here: First, by “learning,” we refer to either training from scratch or further ne-tuning (as commonly done in the LLM context). Second, by “unlearning,” we mean any gradient ascent-based unlearning algorithms. The reason that only gradient ascent-based unlearning can benet from our approach is that, as later described in Section3, Ready2Unlearn modies the model parameters during training toward a state that is particularly amenable to subsequent gradient ascent-based parameter updates. As such, the model is “pre-conditioned” to respond more eciently and reliably, exclusively when gradient ascent is later applied to remove specic data. Ready2Unlearn adopts meta-learning principles, particularly the MAML algorithm [22]. Tradi- tional meta-learning techniques aim tond a model initialization such that performing a small number of gradientdescentsteps from this starting point allows the model to quickly adapt to new tasks. In our setting, we aim to obtain a model “initialization” (the prepared model state) such that, based on this initialization, applying only a few gradientascentsteps (representing unlearning 2 See *GDPR deep dive—how to implement the ‘right to be forgotten’*, available at:https://w.bankinghub.eu/nance- risk/gdpr-deep-dive-implement-right-forgotten?utm_source=chatgpt.com, Fig. 3. ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. 4Hanyu Duan, Yi Yang, Ahmed Abbasi, and Kar Yan Tam operations) leads to a signicant increase in loss for the intended forgotten data, while preserving overall model utility and ensuring strong post-unlearning resilience. To demonstrate preparing models for unlearning readiness at the learning stage is feasible, we test the proposed Ready2Unlearn idea on both vision and language unlearning tasks, including class-wise unlearning and random data unlearning in Section4. In our experiments, we show that with Ready2Unlearn preparation, the model responds much more quickly to unlearning requests compared to models without such preparedness or baseline approaches. Additionally, the model’s overall utility is better preserved, with reduced risk of catastrophic forgetting [15,28,53,63,83]. Furthermore, when attempting to recover the forgotten data by furthernetuning the unlearned model on data with similar distributional characteristics, such as stylistically or semantically related examples, the model exhibits greater resistance to recovery compared to its unprepared counterpart. We hope our work oers fresh perspectives in machine unlearning and additionally inspire future work on incorporating such “looking-ahead” designs in broader application contexts where data are inherently dynamic rather than static, as in personalized recommendation and information retrieval systems [24,46,52,66,80,97]. 2 Related Work 2.1 Machine Unlearning Because of the growing need to eliminate traces of specic data from machine learning models throughout the AI lifecycle, machine unlearning techniques have seen rapid advancement [56,81]. Current research in this area primarily focuses on developing eective unlearning algorithms [11,21,34–36,53,63,96], designing rigorous evaluation protocols [29,38,54,72,93], and identifying and addressing practical, real-world challenges [71], such as the unavailability of user data during the unlearning process [82], which challenges some classic assumptions in traditional unlearning practices. While existing machine unlearning methods generally perform well in typical scenarios, they face critical limitations, such as prolonged unlearning time [20], catastrophic forgetting [15, 28,53,63,83], and vulnerability under more rigorous evaluation conditions, such as susceptibility to jailbreak attacks [94] and the ease with which forgotten data can be recovered [29,38,93]. More importantly, these limitations may become even more pronounced in user-centric applications, such as recommender and information retrieval systems, where models are trained on large volumes of sensitive user interaction data. Although prior work has explored developing new unlearning algorithms or adapting classical approaches to recommendation and information retrieval settings [9,10,16,32,44,49,50,90], existing solutions still leave substantial room for improvement. Our work contributes to this line of research by demonstrating that it is possible to prepare models at the learning stage with future unlearning readiness to further enhance the eciency, reliability, and robustness of existing unlearning methods. We believe this oers a new perspective for enhancing current unlearning practices. 2.2 Training-Time Regularization to Mitigate Memorization Mitigating training data memorization is a reasonable approach to alleviate future unlearning eorts, as models that memorize less are less tied to specic details of data points, making it easier to forget them later [95]. This is typically achieved through regularization techniques applied during model training. Well-known examples include dropout [74], weight decay [40,62], data augmentation [73], and dierentially private learning [1], to name a few. In addition to these general, model-agnostic methods, there are more specialized techniques, such as Goldsh regularization [27], which randomly excludes tokens from loss computation, and NEFTune [33], which adds noise to embedding vectors, both tailored for large language models. At a broader level, the ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. Ready2Unlearn: A Learning-Time Approach for Preparing Models with Future Unlearning Readiness5 Ready2Unlearn method presented in this work can also be positioned within the regularization literature. Furthermore, it adds to this line of research with a novel unlearning-specic regularization technique. Ready2Unlearn is particularly advantageous in machine unlearning scenarios, where existing general regularization techniques often fall short, as they are, by nature, not optimized for unlearning contexts. Thus, we believe this work provides a more targeted solution from this perspective. 2.3 Meta-Learning Another relevant line of research is meta-learning, also known as “learning to learn,” which aims to train models that can rapidly adapt to new tasks with limited data or computational resources [30]. A notable example is model-agnostic meta-learning (MAML) [22], which seeks tond a model initialization such that only a few gradient updates are required for eective adaptation to new tasks. This paradigm has been widely adopted across variouselds, such as domain generalization [42], safeguarding LLMs against adversarial attacks [75], and hyperparameter optimization [2]. Our work draws inspiration from the core idea of MAML, but shifts the objective from fast adaptation to new tasks to fast, reliable machine unlearning. By incorporating a meta-objective at training time, Ready2Unlearn optimizes the model into anunlearning-readystate—a parameter conguration from which future gradient ascent updates (representing unlearning) can proceed in a well-behaved and principled manner. This preparation leads to several desirable properties, including improved unlearning eciency, better retention of overall model capability, and increased resistance to the reintroduction of forgotten information. We are aware of prior work that also applies meta-learning techniques in the context of machine unlearning [31]. However, our approach diers fundamentally in both the timing and the goal of applying meta-learning: their method adopts meta-learning during the unlearning phase, after model deployment, whereas Ready2Unlearn introduces meta- learning at training time, proactively preparing the model before any unlearning request arises. This distinction positions our approach as apreemptivestrategy, marking a departure at the conceptual level from the majority of reactive unlearning methods toward a proactive paradigm. In this sense, we believe that Ready2Unlearn introduces the meta-learning idea to theeld of machine unlearning with a novel use case (i.e., shifting reactive unlearning to a proactive mindset). 2.4 Design Implications for Recommender and Information Retrieval Systems Although the Ready2Unlearn method proposed in this work is generally context-agnostic, we discuss its implications for recommender and information retrieval systems, where machine unlearning has recently attracted increasing attention [9,32,79,91]. For instance, unlearning has a myriad of implications for question answering systems [17] and recommendation engines that are privacy- preserving [85], streaming [70], and/or multi-modal [12,48]. One key idea in Ready2Unlearn is to categorize training data into revocable and stable groups based on their anticipated likelihood of future revocation, thereby improving the eciency of later-stage unlearning. This suggests that data heterogeneity, specically dierences in revocation likelihood across data points, is not a challenge to manage here, but a valuable signal that can be actively exploited in system design. This insight is particularly relevant for recommender systems, where user data naturally exhibit substantial heterogeneity. For example, in personalized recommendation, users often dier in their privacy preferences, disclosure behaviors, and likelihood of requesting data removal [13]. Such dierences can serve as useful signals during training: interaction data from more privacy-sensitive users could be incorporated in ways that allow easier removal when necessary [10], whereas signals from more stable users may play a larger role in shaping the learned model. In this way, user heterogeneity can inform system designs that are not only eective but also better able to support compliance with data protection regulations [64]. ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. 6Hanyu Duan, Yi Yang, Ahmed Abbasi, and Kar Yan Tam Similarly, in information retrieval systems, data heterogeneity also provides useful signals for system design. Documents and knowledge sources often dier in terms of time-sensitivity, reliability, and likelihood of future modication [97]. For example, time-sensitive or user-generated content, such as breaking news updates, forum posts, or other rapidly evolving information, may require frequent correction, updating, or removal [46]. By contrast, well-established reference materials or curated knowledge sources tend to remain relatively stable. Such dierences can be used during system design. Information that is more likely to change can be incorporated in ways that allow it to be easily unlearned from the model when necessary, whereas more stable knowledge can be integrated in a manner that is minimally aected by the updating or removal of other information. In this way, content heterogeneity can help IR systems remain robust and maintain high information integrity, while also enabling adaptability as information evolves over time. Such heterogeneity can also help alleviate biases stemming from familiarity [5], stereotypes [41], and so on. We hope this perspective encourages future work to view data heterogeneity not only as a challenge to manage but also as an opportunity for broader information systems design. 3 Learning with Unlearning Preparedness In this section, we present the proposed Ready2Unlearn method. The notations used in this article are listed in Table1. Table 1. Description of Notations Notations Description DTraining data D f Revocable (forget) data D r Stable (retain) data D rc Recovery data \ 0 Initial model \ % Ready2Unlearn-optimized model \ 0 % Unlearned model \ 00 % Post-recovery model PPreparation operation GAGradient ascent unlearning operation RCRecovery operation #Number of outer-loop optimization steps UAdaptation rate [Learning rate LLoss function _ 1 ,_ 2 , and_ 3 Loss weighting factors ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. Ready2Unlearn: A Learning-Time Approach for Preparing Models with Future Unlearning Readiness7 3.1 Unlearning Process We assume that the model developer has access to a datasetD=D f [D r , which comprises both revocable data (likely to be unlearned in the future, referred to asforget data 3 ,D f ) and stable data (unlikely to be unlearned, referred to asretain data,D r ), and builds a model with weights\ % , where a preparationPhas been applied. Our goal is to designPsuch that\ % performs well on three metrics when future unlearning is triggered:efficiency_metric(\ % ),retention_metric(\ % ), andresistance_metric(\ % ). In this work, by “unlearning,” we refer to the process of applying gradient ascent steps to adjust\ % based on the forget data. Gradient ascent is widely used as an eective unlearning strategy due to its simplicity and its model- and data-agnostic nature [26,28,35,43,58,76,77,88]. Moreover, we assume that the retain data is not accessible during unlearning. We impose this stricter condition because, in many real-world scenarios, access to retain data is often impossible, and retraining the model on this data can be prohibitively costly [14,23,84]. Thus, throughout the paper, unlearning specically refers to a clean process where gradient ascent steps are applied solely to the forget data. 3.2 Problem Formulation and Metrics LetGAdenote the gradient ascent unlearning operation, which maps the prepared model\ % to the unlearned model\ 0 % =GA(\ % ;D f ). LetRCdenote the recovery operation, which furthernetunes the unlearned model\ 0 % on data samplesD rc , similar in style to the forget data, producing the post-recovery model\ 00 % =RC(\ 0 % ;D rc ). Below, we dene three key metrics. –Eciency metric.We say a preparationPleads to ecient unlearning if the model\ % experiences a substantial increase in loss on the forget data after only a few, or even a single, gradient ascent unlearning update. Thus,efficiency_metricis dened as the loss (e.g., classication error) of the unlearned model on the forget data. A higher loss indicates greater unlearning eciency. –Retention metric.We say a preparationPenables strong capability retention if the unlearned model\ 0 % preserves much of its performance on the retain data. Thus, retention_metricis dened as the unlearned model’s performance (e.g., classication accuracy) on the retain data. Higher performance signies stronger retention. –Resistance metric.We say a preparationPequips the model with greater resistance to the inadvertent recovery of erased information if, upon furthernetuning the unlearned model on data similar in style to the forget data, 4 the resulting model is less likely to regain information about the forget data. Thus,resistance_metricis dened as the loss of the post-recovery model\ 00 % on the forget data. A higher loss indicates stronger post-unlearning resilience. 3.3 Unlearning-Ready Training To proactively prepare models during learning towards more ecient and principled unlearning in the future, we introduce Ready2Unlearn, aforward-lookingmethod outlined in Algorithm1. Inspired by meta-learning, this approach prepares the model during learning to beunlearning-ready, ensuring that future unlearning via gradient ascent behaves in a stable and reliable manner. At a high level, we learn the model\ % with unlearning preparedness using adual-loopoptimization 3 Here, in presenting the methodology, we use the terms “forget data” and “retain data” to denote revocable and stable data, respectively, to align with the conventional terminology in the machine unlearning literature. 4 Please refer to Section5.8, where we demonstrate how the similarity can be measured. 5 Although the term “adaptation rate” does not refer to adaptation per se in our setting, we stick to it to maintain consistency with the meta-learning convention. ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. 8Hanyu Duan, Yi Yang, Ahmed Abbasi, and Kar Yan Tam Algorithm 1Ready2Unlearn: Learning with Unlearning Preparedness 1:Input:Initial model parameters\ 0 ; training dataD, composed of forget dataD f and retain dataD r ; recovery dataD rc ; number of outer-loop optimization steps#; adaptation rateU; 5 learning rate[; loss weight coecients_ 1 ,_ 2 , and_ 3 ; and loss functionL. 2:Output:A model with unlearning preparedness\ % . 3:for8=1to#do 4:SampleG f ⇠D f ,G r ⇠D r ,G rc ⇠D rc 5: ˆ \ 81 =\ 81 +Ur \ 81 L ( \ 81 ;G f ) # Inner-loop update mimicking the unlearning. 6:6 0 =r ˆ \ 81 L ⇣ ˆ \ 81 ;G f ⌘ # For improving future unlearning eciency. 7:6 1 =r ˆ \ 81 L ⇣ ˆ \ 81 ;G r ⌘ # To support capability retention after future unlearning. 8:6 2 =r ˆ \ 81 L ⇣ ˆ \ 81 ;G rc ⌘ # To enhance future post-unlearning resilience. 9:SampleG⇠D 10:6 3 =r \ 81 L ( \ 81 ;G ) # For maximizing current model utility. 11:Update\ 8 \ 81 [ ( 6 0 +_ 1 6 1 +_ 2 6 2 +_ 3 6 3 ) # Outer-loop parameter update. 12:end for 13:\ % \ # 14:return\ % structure inspired by MAML [22], comprising aninner-loopgradient update and anouter-loop optimization. 3.3.1 Method Intuition.The key idea behind Ready2Unlearn is tooptimize for the future. Rather than maximizing the model’s immediate performance, we simulate potential unlearning operations that may occur later and optimize the model to be ready for them. Specically, this is achieved by designing the inner-loop gradient update tomimicunlearning actions, representing the “unlearner’s” rst move. In the outer loop, the model parameters are optimized to maximize three desirable properties—eciency, retention, and resistance—against the unlearner’s move simulated in the inner loop. This forward-looking optimization ensures that, when unlearning is eventually triggered, the model will exhibit optimal performance in terms of unlearning eciency, capability retention, and post-unlearning resilience. Essentially, we are preparing the model for future unlearning challenges rather than simply optimizing for the current task. This forward-looking perspective distinguishes our method from conventional approaches. We illustrate this forward-looking idea using conceptual 1D loss landscapes in Figure2. 3.3.2 Unlearning-Ready Objective.LetLdenote the loss function suitable for the task at hand. We now describe the optimization objective of Ready2Unlearn. –Toenhance unlearning eciency(i.e., to achieve fast unlearning), we seek tond a which even a single step of gradient ascent (representing an unlearning step),GA(\), leads to a substantial increase in loss on the forget data. Thus, we consider maximizingL(GA(\);D f ). –Toavoid catastrophic forgeing(i.e., to promote performance retention), we aim to identify a that, after unlearning, the resulting parameters,GA(\), still enable the model to preserve much of its performance on the retain data. Accordingly, we consider minimizing L(GA(\);D r ). ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. Ready2Unlearn: A Learning-Time Approach for Preparing Models with Future Unlearning Readiness9 Fig. 2. An illustration of the forward-looking idea of Ready2Unlearn using conceptual 1D loss landscapes. Panel A:The model state obtained with unlearning preparedness (gray circle) lies adjacent to a steep ascent in the loss landscape with respect to the forget data, such that even a single gradient ascent step can trigger a substantial loss increase, enabling fast and eective unlearning. In contrast, the unprepared model (white circle) lies in a flaer region far from any steep increase, requiring many steps to achieve comparable unlearning.Panel B:The model state optimized by Ready2Unlearn resides in a region where the retain-data loss remains low and stable despite gradient ascent updates on forget data, thereby preserving the model’s utility. The unprepared model resides in a region where unlearning actions (i.e., gradient ascent on forget data) adversely impact performance on retain data, as indicated by a sharp increase in retain loss.Panel C:Around the unprepared model state, the loss landscapes of forget data and recovery data exhibit high similarity. As a result, finetuning on recovery data inadvertently lowers the forget loss as well, undoing the unlearning. In contrast, Ready2Unlearn prepares the model in a region where the recovery loss is low and exhibits a distinct paern from the forget loss, thereby making the model less likely to re-acquire forgoen information during further finetuning on recovery data. –Toachieve more reliable unlearning(i.e., to improve post-unlearning resilience), we aim to nd a that, after unlearning withGA(\)and furthernetuning the unlearned model on recovery data (which shares a similar style to the forget data), the resulting model does not regain signicant information about the forgotten data. We operationalize this by minimizing L(GA(\);D rc ). The rationale behind is to direct the unlearning process so that it removes the most distinctive characteristics of the forget data, rather than merely erasing supercial patterns that may also be present in the recovery data. Otherwise, if the model only unlearns supercial patterns, the loss on the recovery data is likely to increase as well, which is exactly what we aim to prevent by minimizing the aforementioned loss. Put together, our objective is to solve the following optimization problem to obtain the unlearning- ready model\ % : min \ [L(GA(\),D f )+_ 1 ·L(GA(\),D r )+_ 2 ·L(GA(\),D rc )+_ 3 ·L(\;D))], (1) where_ 1 ,_ 2 , and_ 3 are scalar weights for the respective losses. Therst three terms represent the “future objectives,”ensuring that when future unlearning takes place, their respective objectives will be optimized accordingly, reecting the forward-looking nature of Ready2Unlearn. Thenal term serves as the“current objective,”optimizing the model’s current utility prior to any unlearning action being taken. Please refer to Algorithm1for the full optimization procedure. It should be noted that the additional meta-objectives may introduce extra computational overhead compared to standard training. We discuss this in Section5.2. ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. 10Hanyu Duan, Yi Yang, Ahmed Abbasi, and Kar Yan Tam 4 Experiments In this section, we conduct experiments to evaluate Ready2Unlearn on both vision and language tasks across various unlearning settings, including class-wise unlearning and random data unlearn- ing. 4.1 Class-Wise Unlearning in Image Classification 4.1.1 Experiment Setup.We consider image classication tasks using the MNIST [19] and PathM- NIST [86,87] datasets. Each class is treated in turn as the revocable data (i.e., forget data,D f ), prepared at training time for potential future unlearning requests. The remaining classes are treated as stable data (i.e., retain data,D r ), used to evaluate the model’s capability retention after unlearn- ing. 6 Following prior unlearning research [96], we use a convolutional neural network (CNN) as the classier. The model is trained using the negative log-likelihood loss function, and performance is evaluated using classication accuracy. We apply arst-order approximation to compute the meta-gradients (as detailed in lines 6, 7, and 8 in Algorithm1). We use an inner-loop unlearning rateUof1⇥10 5 , an outer-loop learning rate[of2⇥10 4 , and set the loss weights_ 1 and_ 3 to 2 and 4, respectively, 7 during training. In the deployment phase, when unlearning is executed, we apply gradient ascent on the forget data with a step size of1⇥10 5 . 4.1.2 Baseline Methods.We consider several baseline methods that reduce the training imprint of revocable data, enabling more ecient future unlearning.Standard Trainingserves as the basic approach, where the model is trained using stochastic gradient descent (SGD) without distinguish- ing between revocable and stable data.Loss Reweightingreduces the inuence of revocable data during training by assigning it half the loss weight of stable data [37].Noisy Trainingperturbs revocable images by adding standard Gaussian noise scaled by 0.3, which helps to prevent the model from overly memorizing these examples and supports easier unlearning in the future [67,73]. InClipped Training, gradient clipping is applied to revocable data to constrain its inuence on model parameters, which may help prevent overtting and support easier unlearning [60]. Finally, Phased Trainingstarts with training on the full dataset in therst half of the training period and then proceeds to train only on stable data in the second half, allowing the model to initially learn from revocable data without continued exposure, which may ease future unlearning [4,25]. 4.1.3 Results and Discussion.We present the unlearning eciency benchmarking results in Figure 3, which lead to several key observations. First, when the model is trained without consideration for future unlearning, the unlearning process is notably slow. This is evident from the Standard Training baseline, where the accuracy on the forget data remains consistently higher than that of all other methods, which incorporate varying degrees of unlearning preparedness. This underscores the importance of training-time preparation for enabling more ecient unlearning later. Second, among the evaluated methods, Ready2Unlearn exhibits the highest unlearning eciency. We observe that once unlearning begins, the model prepared with Ready2Unlearn immediately undergoes a sharp decline in accuracy on the forget data, reecting a quick response compared to other models. For example, when the model equipped with Ready2Unlearn reaches 50% accuracy, the other models still maintain a relatively high accuracy of around 75% on average. Third, at the moment when unlearning is initiated (marked by the vertical dashed line), the accuracies on the forget data are 6 We additionally investigate in Section5.4the case where the forget and retain sets have overlap, and in Section5.6how the method scales with forget-data diversity. 7 Throughout all experiments in this work, the loss weight parameters are selected based on a held-out validation set, following the principle that the chosen weights should enable the fastest unlearning while ensuring that the training performance does not drop by more than 1% relative to the same model trained without the meta-objectives. For a detailed ablation analysis of the loss weights, please refer to Section5.3. ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. Ready2Unlearn: A Learning-Time Approach for Preparing Models with Future Unlearning Readiness11 Fig. 3. Comparison of unlearning eiciency for MNIST (le) and PathMNIST (right). Each line represents the average forget-data accuracy across all class-wise unlearning seings, where each class is treated as the forget class in turn. All methods are evaluated with the same unlearning rate of1⇥10 5 for a fair comparison. The vertical dashed line marks the moment when unlearning begins. Statistical comparisons reveal that, for both datasets, at the point when the best-performing method (Ready2Unlearn) first falls below random guessing performance on the forget data, its corresponding accuracy is significantly lower than that of the best competing baseline approach (both?<1⇥10 17 , one-tailedC-test, based on 10 independent trials). comparable across all methods, indicating that the superior eciency of Ready2Unlearn does not come at the cost of unduly degraded performance on the forget data prior to unlearning. In other words, the unlearning readiness enabled by our approach does not signicantly compromise the model’s pre-unlearning performance. Overall, the results demonstrate that training a model with unlearning in mind improves the unlearning process at a later stage, with more tailored approaches, such as Ready2Unlearn, oering greater advantages. We also examine how the duration and timing of preparatory steps with Ready2Unlearn aects future unlearning eciency; please see Section 5.1and Section5.7for details. We also evaluate whether Ready2Unlearn equips the model with better capability retention when unlearning is performed. Specically, we examine the model’s accuracy on the retain data after unlearning has driven the forget class accuracy down to the level of random guessing. The results visualized in Figure4reveal that models trained with Ready2Unlearn consistently maintain substantially higher accuracy on retain data compared to those trained without unlearning prepara- tion (i.e., Standard Training). It is important to note that during unlearning, we assume retain data is inaccessible; thus, unlearning is performed solely by applying gradient ascent to the forget data, without any concurrent training on retain data as is done in prior work [31]. Thus, the improved retention of performance is entirely attributed to the training-time preparation, emphasizing the advantage of Ready2Unlearn in handling more complex unlearning scenarios where the retain data is inaccessible, with a forward-looking design. 4.2 Random Data Unlearning in Text Generation 4.2.1 Experiment Setup.We use LLaMA-3.2-1B 8 , LLaMA-3.2-3B 9 , GPT-2, and GPT-2 Medium [65], as the focal language models in our experiments and benchmark unlearning eciency on two widely adopted unlearning corpora: MUSE-Books and MUSE-News [72]. We use their “raw” data splits and adhere to the ocial forget/retain partitions for both datasets. As for evaluating whether Ready2Unlearn improves post-unlearning resilience, we use a separate dataset, the Enron email 8 Hugging Face implementation:https://huggingface.co/meta-llama/Llama-3.2-1B 9 Hugging Face implementation:https://huggingface.co/meta-llama/Llama-3.2-3B ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. 12Hanyu Duan, Yi Yang, Ahmed Abbasi, and Kar Yan Tam Fig. 4. Performance retention for MNIST (le) and PathMNIST (right). Each axis of the radar chart corresponds to a class treated as the forget class. The value on each axis shows the model’s retain accuracy when its forget accuracy reaches random guessing. Statistical tests confirm that, for both datasets and across all forget classes, the accuracy on retain data achieved with preparation is significantly higher than that achieved without preparation (all?<1⇥10 14 , one-tailedC-test, based on 10 independent trials). corpus 10 , which oers an ideal setting for recovery evaluation because of the strong stylistic consistency across email messages. We randomly split the data into three subsets: one forD f , one forD r , and one (D rc ) for furthernetuning the unlearned model to assess potential recovery of forgotten information. Throughout the experiments, by learning, unlearning, and recovering, we refer tonetuning the model with the next-word prediction objective using cross-entropy loss (also used as the performance evaluation metric). For unlearning eciency evaluation, we set the loss scaling factors_ 1 and_ 3 to 2 and 4, respectively. For post-unlearning resilience evaluation, we set _ 2 =3and_ 3 =4. We apply gradient ascent unlearning with a step size of1⇥10 6 . 4.2.2 Baseline Methods.We consider the following baseline methods.Standard TrainingandPhased Trainingare included here as dened in the earlier class-wise unlearning setup. We additionally considerDP-SGD[1,47], whichnetunes the model with dierential privacy using a clipping norm of 0.1. We include two language model-specic techniques:Goldsh[27], which randomly excludes tokens from the loss computation with a probability of 0.25, andNEFTune[33], which injects noise into the embedding vectors with a scaling factor ofU=5. These techniques are applied to revocable data during training to mitigate over-memorization, thereby enabling more ecient unlearning later on. 4.2.3 Results and Discussion.We compare the unlearning eciency of all methods in Figure5. The results indicate that with a very small gradient ascent step size during unlearning (as in this experiment,1⇥10 6 ), achieving sucient forgetting for LLMs typically necessitates a considerable number of optimization steps—meaning that the eects of unlearning are not immediately apparent upon initiation. Fortunately, consistent with thendings from the class-wise unlearning experi- ments, incorporating preparatory steps during training can eectively reduce the time needed to achieve meaningful unlearning later on, without signicantly compromising the model’s perfor- mance prior to unlearning. Notably, models trained with Ready2Unlearn begin to forget earlier 10 Enron Email Dataset:https://w.cs.cmu.edu/~enron/ ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. Ready2Unlearn: A Learning-Time Approach for Preparing Models with Future Unlearning Readiness13 Fig. 5. Unlearning eiciency across models and datasets. Rows correspond to Llama-3.2-1B, Llama-3.2-3B, GPT-2 Medium, and GPT-2 (top to boom), and columns to MUSE-Books and MUSE-News (leto right). The vertical axis shows cross-entropy loss on the forget set; each line corresponds to a training-time strategy. All methods use the same unlearning rate (1⇥10 6 ) for fair comparison. The dashed vertical line marks the start of unlearning. Across all dataset–model combinations, Ready2Unlearn achieves significantly higher forget loss at its first crossing of 20 than the best competing baseline (one-tailedC-test,?<1⇥10 7 , 10 trials). ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. 14Hanyu Duan, Yi Yang, Ahmed Abbasi, and Kar Yan Tam Fig. 6. Loss on forget data across three phases for Llama-3.2-1B (le) and GPT-2 (right). The loss decreases and plateaus during training, rises sharply during unlearning, and then decreases and plateaus again during recovery via further finetuning on stylistically similar data. Statistical analysis shows that, for both models, at Epoch 70, the forget loss achieved with preparation is significantly larger than that achieved without preparation (both?<1⇥10 12 , one-tailedC-test, based on 10 independent trials). Fig. 7. A visual example from the forget set comparing token-level loss aer unlearning a model trained without preparation (Standard Training, top) and with Ready2Unlearn preparation (boom). Color shading indicates relative loss per token, with darker tones representing higher values. In the standard case, loss is spread more uniformly across tokens, suggesting that the model treats all content equally during unlearning rather than prioritizing the removal of more critical information. In contrast, the model prepared by Ready2Unlearn tends to assign higher loss to more distinctive, data-specific tokens, such as login names or passwords (e.g., “pallen”, “ke9davis”), showcasing its focus on unlearning more meaningful, sensitive information rather than generic content. This makes the removed information harder to recover without access to the original forget data. than all baselines and reach noticeable unlearning with fewer gradient ascent steps, demonstrating that our approach generalizes well to transformer-based language models, beyond classic deep neural networks like CNNs. Figure6compares the post-unlearning resilience of models trained with and without Ready2Unlearn preparation. The most notable observation is that, after the loss has plateaued following further netuning the unlearned model on a dataset similar in style to the forget data, the model prepared with Ready2Unlearn consistently maintains a higher loss on the forget data compared to the model without preparation. This indicates that the Ready2Unlearn-prepared model is more resistant to regaining the forgotten information—even when exposed to data with similar characteristics—than its non-prepared counterpart. This resilience arises from the gradient term in line 8 in Algorithm1, which directs the model to unlearn the most distinctive, data-specic features of the forget data, rather than merely suppressing supercial patterns that could easily re-emerge during subsequent (inadvertent)netuning. In contrast, without this preparatory step, the model is more susceptible to relearning those supercial patterns, leading to a lower loss on the forget data afternetuning. See Figure7for an illustration, and Section5.5for additional evidence from the lens of representation analysis. Thus, from this perspective, we believe Ready2Unlearn generates new insights into more targeted machine unlearning [55], where the information to be removed from the model is much more nuanced and selective. ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. Ready2Unlearn: A Learning-Time Approach for Preparing Models with Future Unlearning Readiness15 Fig. 8. Eect of preparation duration on future unlearning eiciency with Ready2Unlearn. This figure visualizes the results for 10 distinct seings, where the number of epochs dedicated to preparation ranges from 2 to 20. The horizontal axis represents the timeline from the start of training to the completion of unlearning. The vertical axis shows the accuracy on the forget set, with lower values indicating more eective forgeing. The dashed line marks the epoch at which unlearning begins (epoch 20). All results are based on experiments conducted using the MNIST dataset. 5 Additional Analyses 5.1 Impact of Preparation Duration on Future Unlearning Eiciency We investigate how the duration of preparatory training with Ready2Unlearn inuences the eciency of subsequent unlearning. Specically, we explore how varying the number of epochs dedicated to preparation during training aects the model’s ability to forget target data eciently when unlearning is later initiated. We consider a total training budget of 20 epochs. For each experimental setting, we allocate thenal"epochs to preparatory training using Ready2Unlearn, while the preceding ( 20" ) epochs follow Standard Training procedures. For example, the setting labeled “6 epochs” represents a scenario in which the model undergoes 14 epochs of Standard Training, followed by 6 epochs of preparation using Ready2Unlearn. The “20 epochs” setting then corresponds to applying Ready2Unlearn throughout the entire training period. The results for 10 distinct settings ("22,4,6,...,20), presented in Figure8, reveal a clear trend that longer preparation with Ready2Unlearn consistently leads to a faster response to future unlearning requests, with the model’s performance on the forget data dropping earlier. This suggests that more extensive integration of Ready2Unlearn during training equips the model with stronger unlearning readiness. 5.2 Computational Cost Analysis Consider that Ready2Unlearn optimizes additional meta-objectives during training, which may introduce extra runtime overhead. Here, we intuitively gauge such incremental computational cost by comparing the training runtimes of Ready2Unlearn and Standard Training. We dene runtime as the duration required for the model to reach converged task performance, and report the relative training time overhead in Figure9, specied as: Relative Overhead (%)= ) R2U ) STD ) STD ⇥100,(2) ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. 16Hanyu Duan, Yi Yang, Ahmed Abbasi, and Kar Yan Tam Fig. 9. Relative training time overhead of Ready2Unlearn compared to Standard Training across dierent model-task configurations. Each configuration is evaluated over 10 independent trials. where) R2U and) STD represent the training times for Ready2Unlearn and Standard Training, re- spectively. The comparison reveals several notable observations. First, as expected, optimizing the extra meta-learning objectives results in increased training time. On average, we observe a 13.7% runtime overhead across model-task congurations. Second, this overhead varies with model architecture, task type, and data modality. Interestingly, smaller models (e.g., CNNs) generally experience a greater relative overhead in comparison to larger ones. 11 Third, wend that the variance in runtime across independent runs tends to be larger for high-capacity models (e.g., Llama-3.2) than for smaller ones (e.g., CNNs). This suggests that extremely large-scale models may be susceptible to unstable optimization dynamics under the added meta-objectives. We highlight this potential risk for practitioners and encourage future work to explore solutions that mitigate such variability and improve stability in large-scale training settings. Overall, there is no free lunch in preparing models in advance for proactive unlearning; nevertheless, we believe that investing a modest amount of extra training time is justied to hedge against potential future uncertainty. Table 2. Classification Performance on Retain and Forget Data with Varying_ 1 Loss Weighting_ 1 Pre-Unlearning Accuracy (%) Post-Unlearning Accuracy (%) Retain Rate (") Forget Rate (") Retain(")Forget(")Retain(")Forget(#) _ 1 =1.096.494.274.854.277.6%42.5% _ 1 =2.098.697.875.656.776.7%42.0% _ 1 =3.098.298.678.665.480.0%33.7% _ 1 =4.097.498.680.567.982.6%31.1% Notes.Results are based on the MNIST classication task. Theretain rateis computed asAcc post (retain data)/Acc pre (retain data), and theforget rateis computed as[Acc pre (forget data)Acc post (forget data)]/Acc pre (forget data). While varying_ 1 , all other weights are held constant. Reported accuracies are taken from the model checkpoint at Epoch 30. 11 This observation seems counterintuitive; we believe a more rigorous and comprehensive assessment that rules out potential confounding factors (such as task complexity) is warranted, and we leave this to future work. ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. Ready2Unlearn: A Learning-Time Approach for Preparing Models with Future Unlearning Readiness17 5.3 Ablations 5.3.1 Varying the Retention Loss Weight_ 1 .We investigate the robustness of Ready2Unlearn by varying the retention loss scaling factor_ 1 from 1.0 to 4.0 (Table2). When setting_ 1 =4.0, once unlearning is triggered, Ready2Unlearn enables the model to maintain strong performance on the retain data, preserving over 80% of its pre-unlearning accuracy. Meanwhile, the accuracy on the forget data drops to 67.9%, indicating a moderate unlearning eciency, with the model exhibiting a roughly 30% reduction in its ability to correctly recognize examples from the forget set. Further reducing_ 1 to 1.0 leads to a signicant improvement in unlearning eciency, with the relative drop in forget-set accuracy (compared to pre-unlearning performance) exceeding 40%, while the model still maintains a considerable portion of its pre-unlearning performance on the retain data (77.6%). It should be noted that while the pre-unlearning performance varies with dierent choices of loss weights, the degree of variation remains modest provided the weighting factors are chosen within a reasonable range. Overall, these observations demonstrate that_ 1 is indeed responsible for governing the trade-obetween unlearning eciency and retention performance, and can be adjusted to appropriately balance the two. 5.3.2 Ablation Results for the Resilience Term.Here, we provide detailed results in Table3, corre- sponding to Figure6. These ablation experiments isolate the eect of the resilience term in Equation 1by comparing models trained with and without this term. The results demonstrate the role of the resilience term in preventing the recovery of forgotten information. Table 3. Ablation Results of the Resilience Term in Ready2Unlearn Resilience Weighting_ 2 Converged Loss on Forget Data Llama-3.2-1BGPT-2 _ 2 =3.05.248.04 _ 2 =04.697.40 Notes.This table shows the converged loss on forget-set data for Llama-3.2-1B and GPT-2 models with (_ 2 =3.0) and without (_ 2 =0) the resilience term. Higher loss indicates stronger resistance to unintentional relearning. 5.4 Eect of Forget-Retain Overlap In practice, real-world forget and retain data may not be perfectly separable into the revocable and stable categories assumed by Ready2Unlearn, resulting in potential overlap between the two. In this section, we investigate this scenario using the MNIST and PathMNIST classication tasks (see setup described in section4). We simulate overlaps by blending varying portions of the retain data into the forget set and report the corresponding retention performance in Figure10. We observe that as a larger portion of retain data is mistakenly included in the forget set, retention performance is slightly compromised after unlearning. This is expected, as the model allocates eort to prepare for “forgetting” data that ultimately should be retained, representing a suboptimal use of model capacity. Despite this, compared with the baseline results in Figure4, the model still performs better than if no proactive preparation had been undertaken at all. This suggests that proactive unlearning preparation is generally benecial and that Ready2Unlearn tolerates reasonable levels of data miscategorization, demonstrating its practical utility. Overall, this investigation indicates that while perfect categorization of forget and retain data may not ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. 18Hanyu Duan, Yi Yang, Ahmed Abbasi, and Kar Yan Tam Fig. 10. Retention performance under varying levels of overlap between forget and retain sets, shown for MNIST (le) and PathMNIST (right). The color gradient indicates the level of mixing, with darker shades corresponding to greater overlap; for example, “20%” means that 20% of the forget set is made up of misclassified retain samples. be feasible in practice, Ready2Unlearn can still provide meaningful advantages over reactive or unprepared training approaches. 5.5 Visualizing the Eect of the Resilience Term From the generation-likelihood perspective (Figure7), we show that the resilience term (the penultimate term in Equation1) encourages the model to forget more distinctive, instance-specic information rather than generic features commonly shared across the broader data distribution. In this section, we provide additional supporting evidence for this eect through the lens of learned representations. Specically, we study two models: one trained with the resilience term active and one in which the resilience term is ablated from the optimization objective. We visualize the representations 12 of the forget and retain data produced by each model using t-SNE [57] in Figure11. Comparing the two subplots, for the model trained without the resilience term (left), the t-SNE embeddings of the forget and retain data exhibit substantial overlap, indicating that the model encodes these two groups along largely similar feature directions. In contrast, when the resilience term is included (right), the embeddings of the forget and retain data are much more separable, with noticeably less overlap. This increased separability suggests that the model has learned to encode the forget data using more distinctive, less broadly shared features, thereby making it harder for subsequentnetuning on data outside the forget set to regain the forgotten information. Overall, this representational analysis provides additional empirical evidence for the “distinctiveness-promoting” role of the resilience term, complementing the token loss analysis (Figure7). We hope this sheds light on developing appropriate loss functions to address key research challenges in targeted machine unlearning. 12 By “representation,” we refer to the aggregated last-layer token hidden states obtained via average pooling. ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. Ready2Unlearn: A Learning-Time Approach for Preparing Models with Future Unlearning Readiness19 Fig. 11. t-SNE visualization of forget and retain data representations for models trained with (right) and without (le) the resilience term. The results are based on GPT-2 and the Enron email dataset. 5.6 Scaling Properties with Respect to Forget-Data Diversity In practical scenarios, the forget set may consist of data that is highly diverse. In this section, we investigate how the performance of Ready2Unlearn scales with respect to the diversity of the forget data. Specically, we extend the class-wise unlearning setup on MNIST. Instead of restricting the forget set to a single class, we gradually increase its diversity by allowing multiple classes to constitute the forget set; a larger number of included classes corresponds to a higher level of diversity. We report in Figure12the post-unlearning retention performance of models trained with Ready2Unlearn under varying levels of forget-data diversity, holding all other experimental settingsxed. The results show that the diversity of the forget set has a statistically signicant eect on the model’s retention capability during unlearning. In general, retention performance decreases as forget-data diversity increases, since preparing the model to forget a broader range of features competes with its ability to preserve performance on the retained data under axed model capacity. However, we also observe that this decline plateaus as diversity continues to increase, suggesting that the trade-ois bounded rather than unbounded. In other words, while more diverse forget data compromise retention during unlearning, the magnitude of the compromise stabilizes. We further expect that higher-capacity models may tolerate greater levels of forget-data diversity before notable degradation appears, a direction that we believe warrants more thorough future investigation. 5.7 When to Intervene in Unlearning Preparation Given that Ready2Unlearn modies the model parameters to a state optimized for future unlearning, it is possible that this “prepared” state can be partially altered or degraded by subsequent training on new tasks or additional epochs. In other words, the benets of preparation may be compromised if further training shifts the model away from the carefully structured parameter conguration intended for ecient unlearning. In this section, we investigate how the timing of the Ready2Unlearn intervention aects later unlearning eciency. We adopt the MNIST digit unlearning setup, where the overall training duration comprises 20 epochs. To study the eect of timing, we apply 4 epochs of Ready2Unlearn optimization at dierent stages of the training process: during therst quarter of training (epochs ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. 20Hanyu Duan, Yi Yang, Ahmed Abbasi, and Kar Yan Tam Fig. 12. Post-unlearning retention performance for Ready2Unlearn-assisted training across varying levels of forget-data diversity. Results are based on the MNIST class-wise unlearning setup, with retention performance averaged over classes. Each experiment is repeated over 10 independent trials. 1-5; Q1), the second quarter (epochs 6-10; Q2), the third quarter (epochs 11-15; Q3), and thenal quarter (epochs 16-20; Q4). In all remaining epochs, standard training is applied. This design simulates varying intervention timings, with earlier interventions exposing the model to a higher likelihood of the prepared state being modied by subsequent training, while later interventions preserve the prepared state closer to the point of unlearning. We evaluate unlearning eciency across these four scenarios and compare them to a baseline in which no unlearning preparation is applied, i.e., standard training throughout. The results, presented in Figure13, indicate that later interventions generally yield better unlearning performance. This is expected, as applying Ready2Unlearn closer to the end of training ensures that the prepared state remains largely intact when unlearning occurs. At the same time, these results highlight a potential risk: the prepared model state can be overwritten or partially degraded by other training updates, which should be noted as a caveat. Importantly, compared to the baseline without any preparation, all intervention timings provide measurable benets. This suggests that performing unlearning preparation is almost always advantageous, regardless of timing, although the eectiveness is maximized when applied later in the training process. Overall, this investigation sheds light on a subtle but important aspect of Ready2Unlearn: the timing of intervention matters, and careful consideration may help maximize the eciency and robustness of later unlearning. 5.8Eect of Similarity Between Forget and Recovery Data on Unintentional Relearning Ready2Unlearn uses a recovery set to prevent unintentional relearning of forgotten data. An interesting question is how the similarity between the recovery and forget data inuences this eect. In this section, we conduct additional experiments to investigate the impact of recovery-forget data similarity on the model’s resistance to unintentional relearning. To quantify similarity, werst encode each sample in both the forget and recovery sets into vectors using the text embedding modelgte-Qwen2-7B-instruct[51]. For each recovery sample, we compute its cosine similarity with every sample in the forget set and take the average similarity as a measure of how close that recovery sample is to the forget set. Based on these average similarities, we divide the recovery set into four quantiles: Q1 contains recovery samples least similar to the forget data, and Q4 contains the most similar ones. ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. Ready2Unlearn: A Learning-Time Approach for Preparing Models with Future Unlearning Readiness21 Fig. 13. Eect of Ready2Unlearn intervention timing on unlearning eiciency. The intervention is applied during dierent quarters of the 20-epoch training process. The results are based on the MNIST task. We then perform Ready2Unlearn using each quantile of recovery data separately and report the post-recovery loss on the forget data in Table4. Dierences in converged loss across quantiles indicate that recovery-forget similarity plays an important role in the resulting model’s resistance to unintentional relearning. In general, higher similarity between the recovery and forget data corresponds to greater resistance, as reected by higher loss values observed when using the most similar quantile (Q4). Overall, we hope this analysis helps shed light on strategies for selecting ecient recovery data for more robust unlearning. Table 4. Post-Recovery Loss on Forget Data Across Recovery Sampleantiles Quantile by Similarity GPT-2 Llama-3.2-1B 17.74.8 27.85.1 38.35.4 48.55.6 Notes.Results are shown across four quantiles of recovery samples, divided by their average similarity to the forget set (Q1 = least similar, Q4 = most similar). Results are reported for GPT-2 and Llama-3.2-1B models, with higher loss indicating stronger resistance to unintentional relearning. 6 Conclusion In this paper, we introduce Ready2Unlearn, a forward-looking approach that proactively prepares neural network models during training to enhance their readiness for future unlearning. By incor- porating preemptive steps into the learning process, our method enables models to unlearn more eciently while preserving much of their utility and exhibiting greater resilience after unlearning. We demonstrate that unlearning should not only be treated as an afterthought but rather as a critical aspect of model lifecycle management that can be proactively addressed through forward-looking ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. 22Hanyu Duan, Yi Yang, Ahmed Abbasi, and Kar Yan Tam training designs. We believe this work oers a new perspective for addressing challenges posed by evolving data governance and privacy demands, particularly in the context of machine unlearning. Acknowledgments K. Y. Tam was supported by the Research Grants Council of Hong Kong [Grant T35-607/23-N]. References [1]Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. 2016. Deep learning with dierential privacy. InProceedings of the 2016 ACM SIGSAC conference on computer and communications security. 308–318. [2]Sungyong Baik, Myungsub Choi, Janghoon Choi, Heewon Kim, and Kyoung Mu Lee. 2020. Meta-learning with adaptive hyperparameters.Advances in neural information processing systems33 (2020), 20755–20765. [3]Clark Barrett, Brad Boyd, Elie Bursztein, Nicholas Carlini, Brad Chen, Jihye Choi, Amrita Roy Chowdhury, Mihai Christodorescu, Anupam Datta, Soheil Feizi, et al.2023. Identifying and mitigating the security risks of generative ai. Foundations and Trends®in Privacy and Security6, 1 (2023), 1–52. [4]Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. InProceedings of the 26th annual international conference on machine learning. 41–48. [5]Marialena Bevilacqua, Kezia Oketch, Ruiyang Qin, Will Stamey, Xinyuan Zhang, Yi Gan, Kai Yang, and Ahmed Abbasi. 2025. When Automated Assessment Meets Automated Content Generation: Examining Text Quality in the Era of GPTs.ACM Trans. Inf. Syst.43, 2 (March 2025), 43:1–43:36.https://doi.org/10.1145/3702639 [6]Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. 2021. Machine unlearning. In2021 IEEE Symposium on Security and Privacy (SP). IEEE, 141–159. [7]Yinzhi Cao and Junfeng Yang. 2015. Towards making systems forget with machine unlearning. In2015 IEEE symposium on security and privacy. IEEE, 463–480. [8]Sungmin Cha, Sungjun Cho, Dasol Hwang, Honglak Lee, Taesup Moon, and Moontae Lee. 2024. Learning to unlearn: Instance-wise unlearning for pre-trained classiers. InProceedings of the AAAI conference on articial intelligence, Vol. 38. 11186–11194. Issue: 10. [9]Chong Chen, Fei Sun, Min Zhang, and Bolin Ding. 2022. Recommendation Unlearning. InProceedings of the ACM Web Conference 2022(Virtual Event, Lyon, France)(W ’22). Association for Computing Machinery, New York, NY, USA, 2768–2777.doi:10.1145/3485447.3511997 [10]Chaochao Chen, Yizhao Zhang, Yuyuan Li, Jun Wang, Lianyong Qi, Xiaolong Xu, Xiaolin Zheng, and Jianwei Yin. 2024. Post-Training Attribute Unlearning in Recommender Systems.ACM Trans. Inf. Syst.43, 1, Article 25 (Dec. 2024), 28 pages.doi:10.1145/3701987 [11]Jiaao Chen and Diyi Yang. 2023. Unlearn What You Want to Forget: Ecient Unlearning for LLMs. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational Linguistics, Singapore, 12041–12052.doi:10.18653/v1/2023.emnlp-main.738 [12]Weixin Chen, Li Chen, Yongxin Ni, and Yuhan Zhao. 2025. Causality-inspired fair representation learning for multimodal recommendation.ACM Transactions on Information Systems43, 6 (2025), 1–29. [13]Ziqian Chen, Fei Sun, Yifan Tang, Haokun Chen, Jinyang Gao, and Bolin Ding. 2023. Studying the Impact of Data Disclosure Mechanism in Recommender Systems via Simulation.ACM Trans. Inf. Syst.41, 3, Article 60 (Feb. 2023), 26 pages.doi:10.1145/3569452 [14]Xinwen Cheng, Zhehao Huang, Wenxin Zhou, Zhengbao He, Ruikai Yang, Yingwen Wu, and Xiaolin Huang. 2024. Remaining-data-free Machine Unlearning by Suppressing Sample Contribution.arXiv preprint arXiv:2402.15109(2024). [15]Dasol Choi, Soora Choi, Eunsun Lee, Jinwoo Seo, and Dongbin Na. 2024. Towards ecient machine unlearning with data augmentation: Guided loss-increasing (gli) to prevent the catastrophic model utility drop. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 93–102. [16]Yizhou Dang, Yuting Liu, Enneng Yang, Guibing Guo, Linying Jiang, Jianzhe Zhao, and Xingwei Wang. 2025. Ecient and Adaptive Recommendation Unlearning: A Guided Filtering Framework to Erase Outdated Preferences.ACM Trans. Inf. Syst.43, 2, Article 49 (Jan. 2025), 25 pages.doi:10.1145/3706633 [17]Sourav Das, Sanjay Chatterji, and Imon Mukherjee. 2026. Augmenting Small Language Model for Better Medical Question Answering through Source Authentication.ACM Trans. Inf. Syst.(Feb. 2026).doi:10.1145/3797887Just Accepted. [18]Thomas De Min, Massimiliano Mancini, Stéphane Lathuilière, Subhankar Roy, and Elisa Ricci. [n. d.]. Unlearning Personal Data from a Single Image.Transactions on Machine Learning Research([n. d.]). ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. Ready2Unlearn: A Learning-Time Approach for Preparing Models with Future Unlearning Readiness23 [19]Li Deng. 2012. The mnist database of handwritten digit images for machine learning research [best of the web].IEEE signal processing magazine29, 6 (2012), 141–142. [20]Ronen Eldan and Mark Russinovich. 2023. Who’s Harry Potter? Approximate Unlearning in LLMs.arXiv preprint arXiv:2310.02238(2023). [21]Chongyu Fan, Jiancheng Liu, Yihua Zhang, Eric Wong, Dennis Wei, and Sijia Liu. 2023. SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classication and Generation. InThe Twelfth International Conference on Learning Representations. [22]Chelsea Finn, Pieter Abbeel, and Sergey Levine. 2017. Model-agnostic meta-learning for fast adaptation of deep networks. InInternational conference on machine learning. PMLR, 1126–1135. [23]Jack Foster, Stefan Schoepf, and Alexandra Brintrup. 2024. Fast machine unlearning without retraining through selective synaptic dampening. InProceedings of the AAAI conference on articial intelligence, Vol. 38. 12043–12051. Issue: 11. [24]Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as Language Processing (RLP): A Unied Pretrain, Personalized Prompt & Predict Paradigm (P5). InProceedings of the 16th ACM Conference on Recommender Systems(Seattle, WA, USA)(RecSys ’22). Association for Computing Machinery, New York, NY, USA, 299–315.doi:10.1145/3523227.3546767 [25]Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. 2013. An empirical investigation of catastrophic forgetting in gradient-based neural networks.arXiv preprint arXiv:1312.6211(2013). [26]Laura Graves, Vineel Nagisetty, and Vijay Ganesh. 2021. Amnesiac machine learning. InProceedings of the AAAI Conference on Articial Intelligence, Vol. 35. 11516–11524. Issue: 13. [27]Abhimanyu Hans, Yuxin Wen, Neel Jain, John Kirchenbauer, Hamid Kazemi, Prajwal Singhania, Siddharth Singh, Gowthami Somepalli, Jonas Geiping, Abhinav Bhatele, and others. 2024. Be like a Goldsh, Don’t Memorize! Mitigating Memorization in Generative LLMs.arXiv preprint arXiv:2406.10209(2024). [28]Tuan Hoang, Santu Rana, Sunil Gupta, and Svetha Venkatesh. 2024. Learn to unlearn for deep neural networks: Minimizing unlearning interference with gradient projection. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 4819–4828. [29]Yihuai Hong, Lei Yu, Haiqin Yang, Shauli Ravfogel, and Mor Geva. 2024. Intrinsic evaluation of unlearning using parametric knowledge traces.arXiv preprint arXiv:2406.11614(2024). [30]Timothy Hospedales, Antreas Antoniou, Paul Micaelli, and Amos Storkey. 2021. Meta-learning in neural networks: A survey.IEEE transactions on pattern analysis and machine intelligence44, 9 (2021), 5149–5169. [31]Mark He Huang, Lin Geng Foo, and Jun Liu. 2024. Learning to unlearn for robust machine unlearning. InEuropean Conference on Computer Vision. Springer, 202–219. [32]Thanh Trung Huynh, Trong Bang Nguyen, Thanh Toan Nguyen, Phi Le Nguyen, Hongzhi Yin, Quoc Viet Hung Nguyen, and Thanh Tam Nguyen. 2025. Certied Unlearning for Federated Recommendation.ACM Trans. Inf. Syst.43, 2, Article 48 (Jan. 2025), 29 pages.doi:10.1145/3706419 [33]Neel Jain, Ping-yeh Chiang, Yuxin Wen, John Kirchenbauer, Hong-Min Chu, Gowthami Somepalli, Brian R Bartoldson, Bhavya Kailkhura, Avi Schwarzschild, Aniruddha Saha, and others. 2023. Neftune: Noisy embeddings improve instructionnetuning.arXiv preprint arXiv:2310.05914(2023). [34]Joel Jang, Dongkeun Yoon, Sohee Yang, Sungmin Cha, Moontae Lee, Lajanugen Logeswaran, and Minjoon Seo. 2023. Knowledge Unlearning for Mitigating Privacy Risks in Language Models. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki (Eds.). Association for Computational Linguistics, Toronto, Canada, 14389–14408.doi:10.18653/v1/2023.acl- long.805 [35]Jiabao Ji, Yujian Liu, Yang Zhang, Gaowen Liu, Ramana Kompella, Sijia Liu, and Shiyu Chang. 2024. Reversing the forget-retain objectives: An ecient llm unlearning framework from logit dierence.Advances in Neural Information Processing Systems37 (2024), 12581–12611. [36]Jinghan Jia, Yihua Zhang, Yimeng Zhang, Jiancheng Liu, Bharat Runwal, James Dienderfer, Bhavya Kailkhura, and Sijia Liu. 2024. SOUL: Unlocking the Power of Second-Order Optimization for LLM Unlearning. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 4276–4292. [37]Alex Kendall, Yarin Gal, and Roberto Cipolla. 2018. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. InProceedings of the IEEE conference on computer vision and pattern recognition. 7482–7491. [38]Yongwoo Kim, Sungmin Cha, and Donghyun Kim. 2025. Are we truly forgetting? a critical re-examination of machine unlearning evaluation protocols.arXiv preprint arXiv:2503.06991(2025). [39]Aravind Krishnan, Siva Reddy, and Marius Mosbach. 2025. Not All Data Are Unlearned Equally.arXiv preprint arXiv:2504.05058(2025). [40]Anders Krogh and John Hertz. 1991. A simple weight decay can improve generalization.Advances in neural information processing systems4 (1991). ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. 24Hanyu Duan, Yi Yang, Ahmed Abbasi, and Kar Yan Tam [41]John P Lalor, Ahmed Abbasi, Kezia Oketch, Yi Yang, and Nicole Forsgren. 2024. Should fairness be a metric or a model? A model-based framework for assessing bias in machine learning pipelines.ACM Transactions on Information Systems 42, 4 (2024), 1–41. [42]Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. 2018. Learning to generalize: Meta-learning for domain generalization. InProceedings of the AAAI conference on articial intelligence, Vol. 32. [43]Guanghao Li, Li Shen, Yan Sun, Yue Hu, Han Hu, and Dacheng Tao. 2023. Subspace based federated unlearning.arXiv preprint arXiv:2302.12448(2023). [44]Meng Li and Haochen Sui. 2025. Causal recommendation via machine unlearning with a few unbiased data. InAAAI 2025 Workshop on Articial Intelligence with Causal Techniques, Vol. 2. [45]Nathaniel Li, Alexander Pan, Anjali Gopal, Summer Yue, Daniel Berrios, Alice Gatti, Justin D Li, Ann-Kathrin Dom- browski, Shashwat Goel, Gabriel Mukobi, et al.2024. The WMDP Benchmark: Measuring and Reducing Malicious Use with Unlearning. InInternational Conference on Machine Learning. PMLR, 28525–28550. [46]Xiaoxi Li, Jiajie Jin, Yujia Zhou, Yuyao Zhang, Peitian Zhang, Yutao Zhu, and Zhicheng Dou. 2025. From Matching to Generation: A Survey on Generative Information Retrieval.ACM Trans. Inf. Syst.43, 3, Article 83 (May 2025), 62 pages. doi:10.1145/3722552 [47]Xuechen Li, Florian Tramer, Percy Liang, and Tatsunori Hashimoto. 2021. Large language models can be strong dierentially private learners.arXiv preprint arXiv:2110.05679(2021). [48]Xiping Li, Aier Yang, Jianghong Ma, Kangzhe Liu, Shanshan Feng, Haijun Zhang, and Yi Zhao. 2026. CPGRec+: A Balance-oriented Framework for Personalized Video Game Recommendations.ACM Trans. Inf. Syst.(Feb. 2026). doi:10.1145/3789264Just Accepted. [49]Yuyuan Li, Chaochao Chen, Xiaolin Zheng, Junlin Liu, and Jun Wang. 2024. Making recommender systems forget: Learning and unlearning for erasable recommendation.Know.-Based Syst.283, C (Jan. 2024), 11 pages.doi:10.1016/j. knosys.2023.111124 [50]Yuyuan Li, Chaochao Chen, Xiaolin Zheng, Weiming Liu, Lingjuan Lyu, Yizhao Zhang, Dan Meng, and Jun Wang. 2023. UltraRE: enhancing RecEraser for recommendation unlearning via error decomposition. InProceedings of the 37th International Conference on Neural Information Processing Systems(New Orleans, LA, USA)(NIPS ’23). Curran Associates Inc., Red Hook, NY, USA, Article 553, 15 pages. [51]Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. 2023. Towards general text embeddings with multi-stage contrastive learning.arXiv preprint arXiv:2308.03281(2023). [52]Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Hao Zhang, Yong Liu, Chuhan Wu, Xiangyang Li, Chenxu Zhu, Huifeng Guo, Yong Yu, Ruiming Tang, and Weinan Zhang. 2025. How Can Recommender Systems Benet from Large Language Models: A Survey.ACM Trans. Inf. Syst.43, 2, Article 28 (Jan. 2025), 47 pages.doi:10.1145/3678004 [53]Shen Lin, Xiaoyu Zhang, Willy Susilo, Xiaofeng Chen, and Jun Liu. 2024. GDR-GMA: Machine Unlearning via Direction-Rectied and Magnitude-Adjusted Gradients. InProceedings of the 32nd ACM International Conference on Multimedia. 9087–9095. [54]Sijia Liu, Yuanshun Yao, Jinghan Jia, Stephen Casper, Nathalie Baracaldo, Peter Hase, Yuguang Yao, Chris Yuhao Liu, Xiaojun Xu, Hang Li, and others. 2025. Rethinking machine unlearning for large language models.Nature Machine Intelligence(2025), 1–14. Publisher: Nature Publishing Group UK London. [55]Yujian Liu, Yang Zhang, Tommi Jaakkola, and Shiyu Chang. 2024. Revisiting Who’s Harry Potter: Towards Targeted Unlearning from a Causal Intervention Perspective. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 8708–8731. [56]Zheyuan Liu, Guangyao Dou, Zhaoxuan Tan, Yijun Tian, and Meng Jiang. 2024. Machine unlearning in generative ai: A survey.arXiv preprint arXiv:2407.20516(2024). [57]Laurens van der Maaten and Georey Hinton. 2008. Visualizing data using t-SNE.Journal of machine learning research 9, Nov (2008), 2579–2605. [58]Pratyush Maini, Zhili Feng, Avi Schwarzschild, Zachary C Lipton, and J Zico Kolter. 2024. Tofu: A task ofctitious unlearning for llms.arXiv preprint arXiv:2401.06121(2024). [59]Anmol Reddy Mekala, Vineeth Dorna, Shreya Dubey, Abhishek Lalwani, David Koleczek, Mukund Rungta, Sadid A Hasan, and Elita A Lobo. 2025. Alternate Preference Optimization for Unlearning Factual Knowledge in Large Language Models. InProceedings of the 31st International Conference on Computational Linguistics. 3732–3752. [60]TomášMikolov et al.2012. Statistical language models based on neural networks.Presentation at Google, Mountain View, 2nd April80, 26 (2012). [61]Michele Miranda, Elena Soa Ruzzetti, Andrea Santilli, Fabio Massimo Zanzotto, Sébastien Bratières, and Emanuele Rodolà. 2024. Preserving privacy in large language models: A survey on current threats and solutions.arXiv preprint arXiv:2408.05212(2024). [62]Andrew Y Ng. 2004. Feature selection, L 1 vs. L 2 regularization, and rotational invariance. InProceedings of the twenty-rst international conference on Machine learning. 78. ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. Ready2Unlearn: A Learning-Time Approach for Preparing Models with Future Unlearning Readiness25 [63]Zhuoyi Peng, Yixuan Tang, and Yi Yang. 2025. Adversarial Mixup Unlearning.arXiv preprint arXiv:2502.10288(2025). [64]Formerly Data Protection. 2018. General data protection regulation (GDPR).Intersoft Consulting, Accessed in October 24, 1 (2018). [65]Alec Radford, Jerey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al.2019. Language models are unsupervised multitask learners.OpenAI blog1, 8 (2019), 9. [66]Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Keshavan, Trung Vu, Lukasz Heidt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Maheswaran Sathiamoorthy. 2023. Recommender systems with generative retrieval. InProceedings of the 37th International Conference on Neural Information Processing Systems (New Orleans, LA, USA)(NIPS ’23). Curran Associates Inc., Red Hook, NY, USA, Article 452, 17 pages. [67]Salah Rifai, Xavier Glorot, Yoshua Bengio, and Pascal Vincent. 2011. Adding noise to the input of a model trained with a regularized objective.arXiv preprint arXiv:1104.3250(2011). [68]Jonas B Sandbrink. 2023. Articial intelligence and biological misuse: Dierentiating risks of language models and biological design tools.arXiv preprint arXiv:2306.13952(2023). [69]Stefan Schoepf, Jack Foster, and Alexandra Brintrup. 2024. Potion: Towards poison unlearning.arXiv preprint arXiv:2406.09173(2024). [70]Chenglei Shen, Yi Zhan, Weijie Yu, Xiao Zhang, and Jun Xu. 2026. Enhancing Bandit Algorithms with LLMs for Time-varying User Preferences in Streaming Recommendations.ACM Trans. Inf. Syst.(Feb. 2026).doi:10.1145/3793543 Just Accepted. [71]Shaofei Shen, Chenhao Zhang, Yawen Zhao, Weitong Chen, Alina Bialkowski, and Miao Xu. 2024. Label-agnostic forgetting: a supervision-free unlearning in deep models. In12th International Conference on Learning Representations, ICLR 2024. International Conference on Learning Representations, ICLR. [72]Weijia Shi, Jaechan Lee, Yangsibo Huang, Sadhika Malladi, Jieyu Zhao, Ari Holtzman, Daogao Liu, Luke Zettlemoyer, Noah A Smith, and Chiyuan Zhang. 2024. Muse: Machine unlearning six-way evaluation for language models.arXiv preprint arXiv:2407.06460(2024). [73]Connor Shorten and Taghi M Khoshgoftaar. 2019. A survey on image data augmentation for deep learning.Journal of big data6, 1 (2019), 1–48. [74]Nitish Srivastava, Georey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: a simple way to prevent neural networks from overtting.The journal of machine learning research15, 1 (2014), 1929–1958. [75]Rishub Tamirisa, Bhrugu Bharathi, Long Phan, Andy Zhou, Alice Gatti, Tarun Suresh, Maxwell Lin, Justin Wang, Rowan Wang, Ron Arel, and others. 2024. Tamper-resistant safeguards for open-weight llms.arXiv preprint arXiv:2408.00761 (2024). [76]Ayush K Tarun, Vikram S Chundawat, Murari Mandal, and Mohan Kankanhalli. 2023. Fast yet eective machine unlearning.IEEE Transactions on Neural Networks and Learning Systems(2023). [77]Anvith Thudi, Gabriel Deza, Varun Chandrasekaran, and Nicolas Papernot. 2022. Unrolling sgd: Understanding factors inuencing machine unlearning. In2022 IEEE 7th European Symposium on Security and Privacy (EuroS&P). IEEE, 303–319. [78]Pengfei Wang, Zongzheng Wei, Heng Qi, Shaohua Wan, Yunming Xiao, Geng Sun, and Qiang Zhang. 2023. Mitigating poor data quality impact with federated unlearning for human-centric metaverse.IEEE Journal on Selected Areas in Communications42, 4 (2023), 832–849. [79]Shang Wang, Tianqing Zhu, Dayong Ye, and Wanlei Zhou. 2025. When machine unlearning meets retrieval-augmented generation (rag): Keep secret or forget knowledge?IEEE Transactions on Dependable and Secure Computing(2025). [80]Wenjie Wang, Xinyu Lin, Fuli Feng, Xiangnan He, and Tat-Seng Chua. 2023. Generative recommendation: Towards next-generation recommender paradigm.arXiv preprint arXiv:2304.03516(2023). [81]Weiqi Wang, Zhiyi Tian, Chenhan Zhang, and Shui Yu. 2024. Machine unlearning: A comprehensive survey.arXiv preprint arXiv:2405.07406(2024). [82]Weiqi Wang, Zhiyi Tian, Chenhan Zhang, and Shui Yu. 2025. Oblivious Unlearning by Learning: Machine Unlearning Without Exposing Erased Data.https://openreview.net/forum?id=wAemQcyWqq [83]Yue Wang, Qizhou Wang, Feng Liu, Wei Huang, Yali Du, Xiaojiang Du, and Bo Han. 2025. GRU: Mitigating the Trade-obetween Unlearning and Retention for Large Language Models.arXiv preprint arXiv:2503.09117(2025). [84]Yaxuan Wang, Jiaheng Wei, Chris Yuhao Liu, Jinlong Pang, Quan Liu, Ankit Parag Shah, Yujia Bao, Yang Liu, and Wei Wei. 2024. LLM Unlearning via Loss Adjustment with Only Forget Data.arXiv preprint arXiv:2410.11143(2024). [85]Zihong Xian, Enyue Yang, Weike Pan, and Zhong Ming. 2026. FedHoG: Federated Homogeneous Graph Neural Network for Privacy-Preserving Recommendation.ACM Trans. Inf. Syst.(March 2026).doi:10.1145/3787468Just Accepted. [86]Jiancheng Yang, Rui Shi, and Bingbing Ni. 2021. MedMNIST Classication Decathlon: A Lightweight AutoML Benchmark for Medical Image Analysis. InIEEE 18th International Symposium on Biomedical Imaging (ISBI). 191–195. ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026. 26Hanyu Duan, Yi Yang, Ahmed Abbasi, and Kar Yan Tam [87]Jiancheng Yang, Rui Shi, Donglai Wei, Zequan Liu, Lin Zhao, Bilian Ke, Hanspeter Pster, and Bingbing Ni. 2023. MedMNIST v2-A large-scale lightweight benchmark for 2D and 3D biomedical image classication.Scientic Data10, 1 (2023), 41. [88]Yuanshun Yao, Xiaojun Xu, and Yang Liu. 2023. Large language model unlearning.arXiv preprint arXiv:2310.10683 (2023). [89]Charles Yu, Sullam Jeoung, Anish Kasi, Pengfei Yu, and Heng Ji. 2023. Unlearning Bias in Language Models by Partitioning Gradients. InFindings of the Association for Computational Linguistics: ACL 2023, Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki (Eds.). Association for Computational Linguistics, Toronto, Canada, 6032–6048. doi:10.18653/v1/2023.ndings-acl.375 [90]Wei Yuan, Hongzhi Yin, Fangzhao Wu, Shijie Zhang, Tieke He, and Hao Wang. 2023. Federated unlearning for on-device recommendation. InProceedings of the sixteenth ACM international conference on web search and data mining. 393–401. [91]Peng-Fei Zhang, Guangdong Bai, Zi Huang, and Xin-Shun Xu. 2022. Machine Unlearning for Image Retrieval: A Generative Scrubbing Approach. InProceedings of the 30th ACM International Conference on Multimedia(Lisboa, Portugal)(M ’22). Association for Computing Machinery, New York, NY, USA, 237–245.doi:10.1145/3503161.3548378 [92]Ruiqi Zhang, Licong Lin, Yu Bai, and Song Mei. 2024. Negative preference optimization: From catastrophic collapse to eective unlearning.arXiv preprint arXiv:2404.05868(2024). [93]Zhiwei Zhang, Fali Wang, Xiaomin Li, Zongyu Wu, Xianfeng Tang, Hui Liu, Qi He, Wenpeng Yin, and Suhang Wang. 2024. Does your LLM truly unlearn? An embarrassingly simple approach to recover unlearned knowledge.arXiv preprint arXiv:2410.16454(2024). [94]Chenxu Zhao, Wei Qian, Yangyi Li, Aobo Chen, and Mengdi Huai. 2024. Rethinking Adversarial Robustness in the Context of the Right to be Forgotten. InInternational Conference on Machine Learning. PMLR, 60927–60939. [95]Kairan Zhao, Meghdad Kurmanji, George-Octavian Bărbulescu, Eleni Triantallou, and Peter Triantallou. 2024. What makes unlearning hard and what to do about it.Advances in Neural Information Processing Systems37 (2024), 12293–12333. [96]Juexiao Zhou, Haoyang Li, Xingyu Liao, Bin Zhang, Wenjia He, Zhongxiao Li, Longxi Zhou, and Xin Gao. 2023. A unied method to revoke the private data of patients in intelligent healthcare with audit to forget.Nature Communications14, 1 (2023), 6255. Publisher: Nature Publishing Group UK London. [97]Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan Liu, Chenlong Deng, Haonan Chen, Zheng Liu, Zhicheng Dou, and Ji-Rong Wen. 2025. Large Language Models for Information Retrieval: A Survey.ACM Trans. Inf. Syst.44, 1, Article 12 (Nov. 2025), 54 pages.doi:10.1145/3748304 ACM Trans. Inf. Syst., Vol. 1, No. 1, Article . Publication date: March 2026.