Machine Unlearning

Vision-Language Models (VLMs) exhibit remarkable zero-shot generalization, yet they often encode unwanted or hazardous stylistic domains such as idealized textbook diagrams in medical AI or cartoon vehicles in autonomous driving. Approximate Domain Unlearning (ADU) aims to selectively erase a model's recognition of a target visual domain while preserving accuracy on the remaining domains. However, existing ADU methods operate under a flawed closed-vocabulary assumption: they evaluate unlearning solely on the specific object classes seen during the unlearning fine-tuning phase. Consequently, these methods do not unlearn the domain itself; they merely overfit to seen class-domain pairs, leaving the domain easily recognizable for unseen classes and providing a false sense of removal. We argue that true domain erasure must be class-agnostic. To address this, we formalize Open-Vocabulary Domain Unlearning (OVDU), a rigorous protocol that mandates domain forgetting must transfer to held-out classes. To solve the OVDU challenge, we propose a surgical parameter-editing framework. First, a Fisher Information mask isolates domain-sensitive weights, mathematically protecting foundational zero-shot generalization. Second, our Targeted Manifold Scattering (TMS) objective uses preference-based mining to locally scatter the forget domain's stylistic geometry. Evaluated across PACS, OfficeHome, and DomainNet, our method vastly improves open-vocabulary generalization over existing baselines. Crucially, it delivers exceptional sample efficiency, outperforming peak 8-shot baseline results with only 4 shots.

Machine unlearning aims to remove targeted information while preserving a model's other abilities. In realistic settings, such as privacy requests under the EU GDPR, the target may be narrow, for example information associated with a single person. Behavioral forgetting alone may be insufficient, motivating interventions directly on internal representations. However, standard mechanistic-interpretability extractors are poorly selective for such targets. We identify an energy bias in reconstruction-based extraction, which favors dominant background structure over low-energy target-specific components. We introduce SCALPEL, a contrastive sparse autoencoder designed to learn more selective forget features. We show theoretically that contrastive training promotes target-selective features and that our selection score controls expected background knowledge perturbation. We validate SCALPEL experimentally on TOFU across Qwen, Llama, and Gemma, where it substantially improves over NMF and standard SAE interventions and is competitive with Gradient Difference and RMU, bridging mechanistic interpretability and fine-grained unlearning.

In this paper, a method for achieving exact unlearning for Gibbs supervised learning algorithms is proposed using a variational formulation inspired by empirical risk minimization subject to relative entropy regularization (ERM-RER). Such a method consists of maximizing the expected empirical risk over the dataset to be unlearned subject to a regularization by relative entropy with respect to the original algorithm. The optimization variable is a probability measure on the models; and the solution is another Gibbs probability measure that represents a new Gibbs supervised learning algorithm. The method guarantees exact unlearning in the sense that the new Gibbs algorithm coincides in distribution with the algorithm that would have been obtained by retraining from scratch on the dataset to be retained. As a byproduct, a framework for reweighting data points in ERM-RER by strategically choosing both the reference measure and the regularization factor is obtained. In this framework, exact unlearning is the special case in which zero-weight is assigned to the contribution of the data points to be unlearned. More generally, depending on the choice of certain parameters, data points can be up-weighted or down-weighted in ERM-RER problems for particular purposes, e.g., controlling the generalization error of Gibbs algorithms. This paves the way for new constructive or adversarial views on classical reweighting data points in ERM-RER.

Humanoid control, leveraging human demonstrations, has achieved diverse, agile, and natural locomotion behaviors through reinforcement learning (RL). While this paradigm has yielded remarkable performance in physical humanoid control, how to eliminate specific motions from learned policies remains insufficiently explored. Addressing this issue is motivated by pressing safety and privacy concerns: the removal of malicious, poisoned, or suboptimal motions, as well as copyright-protected motions subject to the right to be forgotten under regulations such as the GDPR, is of critical importance. To this end, we propose {ForgetMimic}, the first motion-level unlearning method designed specifically for physical-world humanoid control. The core idea of ForgetMimic is as follows: given a policy $π_θ$ trained on $N$ motions, our method degrades performance on a target subset of $K$ motions while preserving the effectiveness of the remaining $N-K$ motions. Furthermore, we identify and resolve two key training mechanisms in robot control that lead to unlearning failure. We conduct extensive experiments on the Unitree G1 and H2 humanoid robots across 12 motions, including Dance, Fight, Flip, and others. Experimental results demonstrate that ForgetMimic effectively eliminates memory of designated motions while maintaining the normal operation of all other motions.

Concept erasure methods that operate via linear projection assume that features occupy separable subspaces. We show this assumption fails under dense superposition: when two features are forced into an antipodal pair sharing a single subspace, state-of-the-art linear erasure destroys both, not just the target. Networks trained with gradient descent instead solve this problem non-linearly, but not uniformly: they converge to one of two distinct circuit-level solutions depending on initialization, which we call mirror and shadow solutions. We map this bifurcation as a function of feature entanglement, show it reflects a stable attractor structure rather than an artifact of our setup, and use targeted causal interventions to demonstrate that both solutions leave a substantial, measurable trace of the erased feature's representation intact, recoverable through a single scalar patch rather than requiring any further training. This mirrors a failure mode recently observed empirically in LLM unlearning, where suppression rather than deletion allows forgotten knowledge to resurface; our results offer a mechanistic, causally-validated account of why that failure mode occurs.

Recent zero-shot text-to-speech (ZS-TTS) systems can reproduce a speaker's voice with high fidelity from only a few seconds of reference speech, raising concerns over unauthorized voice cloning and impersonation. Speaker identity unlearning has recently emerged as an approach to selectively suppress this capability for speakers who opt out while preserving synthesis capability for other speakers. Although existing approaches reduce speaker similarity, preventing re-identification often faces severe degradation of speech quality. Motivated by this observation, we propose GUARD, a lightweight speaker identity unlearning framework that combines a learned speaker gate with speaker-agnostic activation steering on a frozen TTS backbone. The steering vectors are optimized using group-relative reward optimization to shift outputs from forget speakers toward population-level impostor similarity while preserving intelligibility and speech naturalness. On CosyVoice2, GUARD reduces forget-speaker similarity from 0.541 to 0.103 and re-identification accuracy in a 150-speaker gallery from 73.5% to 0.5%, while preserving retain-speaker reproduction. The results demonstrate that similarity reduction alone may not fully characterize successful speaker identity unlearning and highlight re-identification as a complementary criterion for its evaluation.

Unlearning ensures LLM compliance by removing the influence of private or copyrighted training data. However, since LLM models typically undergo post-training compression, like quantization, in practical deployment, it has been observed that the unlearning effect can be substantially weakened, with the forgetting behavior degrading more severely than that of model utility. This paper proposes a quantization-robust unlearning framework that makes forgetting robust to quantization while maintaining overall model utility. We analyze this gap through the lens of loss landscape. Specifically, our analysis reveals a curvature-based criteria that pinpoints sensitive weights in the unlearned model that leads to both non-robust forgetting and reduced utility. We therefore propose sensitivity-guided noisy regularization, which is applied on the sensitive parameters to steer the model convergence towards a smoother minima of uniformly low forget and retain losses. Balancing unlearning and utility, we further propose forget-critical optimization, which updates only forget-critical layers, preserving most of the network to retain useful knowledge. Extensive experiments on the MUSE and TOFU benchmarks across multiple LLM unlearning algorithms show that our approach achieves substantially more quantization-resilient forgetting while maintaining utility. Code is available at https://github.com/jialuw0830/QRUnlearn.

A face recognition model links two images of a person recorded on separate occasions when their embedding similarity exceeds an operating threshold. We consider making chosen identities unlinkable across separate occasions while the model remains in service for the rest of the population. Deleting their images and retraining does not achieve this, since the model recognises identities never observed in training. Therefore, the embedding space must be altered against these identities, the process of which we call open-set adversarial forgetting. We propose three loss functions, one that disperses an identity's embeddings from their centroid, and two that map each image onto its own near-orthogonal target, learnt with the classifier head or fixed in advance as an almost-orthonormal frame. Each is fine-tuned alongside the classification objective on a subset of each identity's images. We evaluate them against four methods from prior work in verification and identification, at two forget scales and three backbones. Every loss acting on the embedding geometry makes the forget identities nearly unidentifiable. The orthonormal frame alone achieves strong forgetting, which holds wherever an image of that subset enters the comparison and leaves distinct forget identities unlinkable. It also surpasses a concurrent unsupervised method at a higher retain rate.

Diffusion data-point unlearning is typically evaluated immediately after each deletion, even though subsequent requests may repeatedly update the same model. We identify sequential reappearance, a failure mode in which an instance that is initially judged to be forgotten later returns to the memorized regime without reuse of the deleted data or adversarial fine-tuning. To capture this behavior, we introduce a target-level evaluation protocol that tracks whether each target is forgotten immediately, remains forgotten at the end of the sequence, or reappears during subsequent deletions. We further find that targets that later reappear exhibit sharper local denoising-loss geometry after deletion than targets that remain forgotten.

Recent advances in large reasoning models (LRMs) have made machine unlearning more challenging, as protected facts or unsafe rationales may surface in intermediate chain-of-thought (CoT) traces before the final answer is produced. Existing unlearning objectives typically suppress the target content or redirect internal representations, but they never specify how the post-forgetting trajectory should continue, which can lead to hallucinated substitutes, malformed boundaries, or repetitive outputs. We argue that LRM unlearning should instead learn a natural forgetting trajectory: a coherent non-disclosing CoT followed by a stable refusal-style answer that replace the original disclosure. To this end, we propose Guided Answer-Reasoning Distillation (GUARD), which converts model-generated unsafe disclosures into safe-exit trajectories, aligns a frozen LRM via guidance tokens, and distills the guided behavior into model parameters. To address the lack of metrics for replacement quality beyond leakage, we further introduce Natural Forgetting Reasoning Score (NFRS), which captures structural stability, fluency, and unsupported substitutes in forgotten outputs. Extensive experiments on R-TOFU and a STAR-1-derived harmful-intent setting show that GUARD substantially reduces unsafe and privacy disclosures across two widely adopted distilled LRMs while preserving reasoning utility.

Undesired information such as harmful content and private data propagates through Multilingual Large Language Models (LLMs) via direct training and indirect cross-linguistic spread. Multilingual Machine Unlearning (MMU) aims to remove such information, yet its evaluation remains underexplored, leaving unclear whether unlearning truly eliminates target knowledge across all languages. To bridge this gap, we introduce $μ^2$-Bench, an MMU benchmark that simulates the full pipeline of memorization, unlearning, and evaluation across diverse languages. It 1) spans a broad set of languages, 2) evaluates on both training and hold-out languages, and 3) assesses knowledge as dispersed across multiple languages. We show that successful MMU requires methods that reflect multilingual characteristics, and conduct analysis to provide deeper insights into MMU.

Pre-trained vision-language models (VLMs) exhibit strong cross-domain recognition performance even without additional training. However, this robustness can also preserve undesirable domain-specific behavior, as domain-related and semantic information often remain entangled within the learned representation space, making selective domain unlearning challenging. Existing approaches typically address this problem through latent-space disentanglement and prompt- or feature-level interventions, without directly attributing and attenuating individual patch-token contributions. However, here we suggest that rather than uniformly suppressing the full representation, it may be more effective to exploit the spatial structure of vision transformers to localize and suppress patch regions that contribute disproportionately to forget-domain prediction. Patches that strongly influence forget-domain prediction may not be equally important for semantic recognition, suggesting that forgetting should be guided according to the domain contribution of different visual regions. Specifically, we propose a two-stage patch-selective framework that first estimates patch-level domain sensitivity and then selectively attenuates patches whose contribution to forget-domain prediction is stronger than their semantic utility. We evaluate our framework on Office-Home, Mini DomainNet, and DomainNet. Experimental results demonstrate improved forgetting-retention tradeoffs compared to prior methods while improving retained-domain recognition by up to 3.8%. Additional evaluations under visually overlapping and unseen-domain settings further demonstrate improved robustness under distribution shift.

Large Language Model (LLM) unlearning is essential for removing sensitive or copyrighted knowledge while preserving general utility. Existing methods often leave residual knowledge in intermediate representations, which can still be recovered. To address this, we propose Cascade, a hierarchical recoverability control framework that minimizes the internal identifiability of target knowledge. Cascade combines three complementary controls: path-level routing to suppress privacy-associated activation routes, representation-level compression to reduce geometric separability, and decoding-level intervention to limit residual recovery. Experiments on TOFU, MUSE-News, and WMDP, including robustness tests with query reformulation and extraction-style prompts, show that Cascade effectively reduces recoverability while maintaining stable model utility.

Machine unlearning aims to remove specific knowledge from a trained large language model (LLM) without retraining from scratch. Existing methods modify model weights via gradient ascent and its advances. While effective on certain benchmarks, these weight-based approaches exhibit a sharp forget-utility trade-off, where stronger forgetting of target knowledge can degrade model utility, and unlearned knowledge may reappear under post-unlearning fine-tuning or prompt attacks. We propose ARIA (autoencoder-gated inference-time unlearning), a test-time unlearning method that leaves model weights intact and gates access to unwanted knowledge only when generation enters a forget-related state. ARIA uses sparse autoencoder (SAE) latents to train a lightweight linear detector, then applies an interpretable intervention on triggered states with negligible test-time overhead. Empirical evaluations on TOFU, R-TOFU, and WMDP show that ARIA improves the forget-retain trade-off over weight-based baselines across both a thinking model (DeepSeek-R1-Distilled-Qwen-1.5B) and an instruction model (Gemma-3-1B-it), e.g., reducing WMDP-cyber forget-set accuracy significantly while keeping MMLU within 1% of the pre-unlearning model. We further introduce three post-unlearning adversarial attacks targeting weight-space and decoding-space recovery, and find that ARIA remains robust under all three, with forgetting changing by less than 1% under attack. A feature-level case study leveraging the interpretability of ARIA suggests that some retain degradation may reflect response styles underlying the unlearning data rather than leakage of the targeted knowledge itself, highlighting a potential source of bias in unlearning task construction.

As deep learning models become fundamental to modern healthcare, the "Right to be Forgotten" mandated by privacy regulations like GDPR and HIPAA necessitates effective machine unlearning (MU) to remove sensitive patient data from trained models. However, existing MU techniques often struggle with a fundamental "privacy-efficiency-utility" (PEU) trilemma, particularly in medical scenarios where data is frequently characterized by severe class imbalance and long-tailed distributions. In such cases, standard unlearning methods can fail to protect key clinical knowledge or mistakenly delete features essential for diagnosing rare conditions due to the gradient dominance of majority classes. To address these challenges, we propose GRIN+, a novel machine unlearning framework designed for fast and precise data erasure in imbalanced medical scenarios. GRIN+ decouples unlearning-specific knowledge from generalized representations at the parameter level by analyzing the gradient contributions of both "forget" and "retain" sets. It introduces a class-adaptive influence scoring mechanism to rectify gradient dominance and employs a direction-constrained update strategy to prevent the unintended erosion of vital clinical knowledge. Comprehensive benchmarking across multiple medical datasets, including skin cancer (ISIC), brain tumor (MRI), and breast ultrasound (BUSI), demonstrates that GRIN+ achieves an optimal balance of the PEU trilemma. Experimental results show that GRIN+ maintains high diagnostic accuracy and robust privacy while significantly enhancing runtime efficiency compared to existing baselines. We open-source the GRIN+ code and benchmarks to support further research.

The increasing adoption of machine learning in network and distributed security systems has created an urgent need for mechanisms that can selectively and efficiently remove the influence of specific training data to eliminate compromised or adversarial data points from production models. Privacy regulations such as GDPR's right to be forgotten also pose similar requirements. However, existing approximate unlearning techniques remain computationally prohibitive for deployment in real-world security systems, as they require repeated expensive Hessian-inverse-vector computations for each data point removal, creating a bottleneck when processing multiple related requests in scenarios such as intrusion detection systems, spam filters, and threat intelligence platforms. Thus, we introduce a computationally efficient unlearning framework that identifies correlated data points in the training set and applies a theoretically derived closed-form parameter update rule, achieving an $82\times$ wall-clock speedup over standard influence function unlearning while preserving model utility with a $10^{-2}$ improvement in accuracy over state-of-the-art baselines. Our method establishes theoretical guarantees and ensures numerical stability through Hessian damping. Our evaluation across seven diverse dataset architecture combinations, including large-scale CIFAR-100 with ResNet-50, demonstrates superior forgetting effectiveness, with membership inference attack success rates of 0.660 and tug-of-war scores of 0.950.

We evaluate Deep Perturbation Learning (DPL), which perturbs training images and labels along influence-derived directions, in three roles in which prior work has positioned it for machine unlearning: a direct deletion signal (the strongest claim), a utility-preserving regularizer, and a warm start for adversarial unlearning. Evidence for the weaker roles has been used to support the stronger one, so we test each role separately under a matched protocol with exact-seed retraining baselines. An audit of the public implementation identifies two correctness issues: image directions are computed on augmented, normalized tensors but applied to raw images, and the label perturbation falls below float32 resolution, leaving labels unchanged. After correcting the image-perturbation pipeline, DPL fails the direct-deletion criterion on CIFAR-10/ResNet-18 in all three paired seeds. Its utility effects are inconsistent in sign across seeds, and once direction-computation time is counted it underperforms simple warm-start baselines. A one-seed Tiny ImageNet check likewise does not favor DPL as a regularizer or warm start; preprocessing inconsistencies in the released code make the direct comparison there inconclusive. These results cover random instance deletion only and do not rule out influence-based methods in other deletion regimes. We release a role-matched evaluation protocol and an audit checklist for perturbation-based deletion claims.

Existing evaluations of concept unlearning in text-to-image (T2I) diffusion models primarily rely on attack success rates obtained through automated adversarial prompt search. However, these metrics provide only empirical evidence over a finite set of queries and leave residual leakage over the broader prompt space largely unquantified. This limitation can lead to overestimating unlearning effectiveness and underestimating safety risks. To address this gap, we introduce a novel certification framework for T2I concept unlearning that provides high-confidence guarantees with bounded error on residual concept leakage. Our approach combines statistical certification with worst-case analysis along concept-relevant embedding directions to derive explicit upper bounds on leakage probability under user-specified confidence levels. We evaluate our framework across three major concept categories namely NSFW content, artistic styles, and celebrity identities, and six state-of-the-art unlearning methods. Certified leakage bounds consistently exceed standard attack success rates by 16.2%, uncovering substantial residual risks missed by existing evaluation protocols. Crucially, our results demonstrate that empirical attack-based evaluations can significantly underestimate residual leakage and establish certification as a necessary complement for reliable auditing of concept unlearning in T2I diffusion models.

An unlearning audit reads its verdict off numbers that an unlearned model and its retrained reference each publish, and both also ship batch-normalization statistics that no gradient step wrote and no release records. Refitting them on kept data at bit-identical weights moves 47 of 221 released checkpoints past the spread their own release's seeds show, several inside a method whose average does not move: what moves is the checkpoint's property, not its method's. What does the moving is not the removed data surviving in the state: exchanging kept records for removed ones inside a fixed fitting pool moves a published cell by almost nothing, while how far a checkpoint's shipped state has drifted from any refit does track it. The consequence for a published decision is real but narrow: twelve verdicts cross, four clear a measured recalibration budget, two clear it on every replicate, and a population we trained and sited near its own criterion yields none. A release should therefore name the fitting convention beside the number, on the batch-normalized vision models where this channel exists.

Large Language Models (LLMs) can memorize and reproduce sensitive, copyrighted, or otherwise undesirable training content, creating privacy, safety, and regulatory concerns. Machine unlearning offers a practical alternative to full retraining, but many existing methods apply broad or fixed parameter updates that can degrade utility and remain brittle under deployment changes such as post-training quantization, where forgotten knowledge may partially re-emerge. We propose Forgetting Only What Matters via Unlearning Layers (FOM-UL), a layer-level unlearning framework that selects transformer layers using a forget-to-retain significance score. This score identifies layers with high influence on the forget set and low sensitivity to the retain set, allowing FOM-UL to concentrate updates where they are most effective while leaving most of the model unchanged. This targeted update strategy improves the forgetting-utility trade-off and provides an empirical path toward quantization-resilient unlearning by reducing the chance that small, diffuse updates are erased by low-bit rounding. Across TOFU, KnowUnDo, and MUSE-style evaluations, FOM-UL reduces residual memorization compared with strong GA, NPO, KLD, SURE, ReLearn, and LUNAR-based baselines while preserving retain-set utility close to the vanilla model. Under 8-bit and 4-bit post-training quantization, FOM-UL maintains stronger memorization suppression and utility preservation than competing methods, and adversarial prompt evaluations show lower recovery of forgotten content. Overall, FOM-UL provides an efficient unlearning strategy that improves targeted forgetting, utility preservation, and deployment robustness without claiming formal guarantees of erasure.

Approximate machine unlearning aims to remove the influence of specific training data from a trained model without retraining from scratch. We identify a previously undocumented confound in how unlearning is evaluated on BatchNorm-based architectures: a single forward pass over retain data, an operation that modifies no weight, can deterministically rewrite the model's normalization state and reverse the apparent surface-metric forgetting. We formalize this operation as a weight-preserving fixed-point operator and prove that any pre-versus-post gap it induces is provably attributable to BN running statistics rather than to any modification the unlearning method made to the weights. This attribution claim cleanly separates measurement failure (BN artifact) from encoder failure (residual weight-encoded information, recently documented in concurrent work), and the same operator framework yields a unique decomposition of linear-probe elevation into BN-measurement-bias and encoder-geometry components. Empirically, the artifact reverses headline forget accuracy by up to 78 pp across nine evaluated methods on standard benchmarks; an attacker with as few as 10 unlabeled images recovers most of the masked accuracy; and a strict GroupNorm control reduces the artifact to zero across all methods. The tested membership-inference attacks change little under recalibration, locating the observed evaluation failure in forget accuracy and linear probing.

Graph unlearning removes the influence of designated training data from a trained graph model without retraining from scratch. However, existing methods suffer a sharp drop in model utility under large deletion ratios (mass deletion), a phenomenon we refer to as catastrophic unlearning. We find that a key cause is the uniform treatment of all deleted samples, which is particularly damaging in graph learning: structural dependencies cause different nodes to play vastly different roles in the learned model, yet existing methods apply the same forgetting operation to the entire forget set. Based on this insight, we propose CUNO, a curriculum-based graph unlearning framework that removes the forget set progressively, ordering samples by their estimated unlearning difficulty across multiple stages. CUNO further employs a distribution-level negative preference optimization (NPO) objective at each curriculum stage that steers the model away from its original behavior on the current forget subset while preserving retained performance. Our theoretical analysis shows that the curriculum design is most beneficial when the forget set spans a wide range of unlearning difficulty, a condition naturally satisfied under mass deletion. Comprehensive experiments confirm that CUNO consistently mitigates catastrophic unlearning: at 20% deletion, it retains 74% of the original utility compared to 26-53% for existing methods, and maintains more than half the original utility even at 50% deletion. Our code is publicly available at https://anonymous.4open.science/r/cuno-D4FF.

Machine unlearning aims to remove the influence of designated training data while preserving model utility, but its behavior on tabular data remains underexplored. This gap is important because tabular prediction is widely used in high-stakes domains and is increasingly adapted to language models through record serialization and schema-aware prompting. We identify a key challenge that distinguishes tabular unlearning from unlearning in free-form text or other modalities: schema-induced forget-retain overlap. In serialized tabular data, records share fixed column-name/value slots, similar attribute ranges, and common output spaces. Consequently, a forget row may have nearby retain rows that rely on the same high-signal attributes, causing retain preservation to oppose the update required for forgetting. Motivated by this failure mode, we propose Conflict-Aware Unlearning (CAU), a schema-aware approach that reduces forget-retain interference by relaxing preservation constraints on retained rows that most conflict with the forget set. Across sample-level and feature-level unlearning on clinical and non-medical tabular tasks, CAU more closely matches a retraining oracle while maintaining predictive utility and retain-region behavior. Our results show that reliable tabular LLM unlearning depends not only on the forgetting objective, but also on how retain constraints are constructed.

Existing unlearning approaches typically rely on post hoc weight adaptation or distillation, leading to duplicated memory costs, degraded generalization, and limited scalability. In this work, we introduce ERASE, Erasure via Reconstructive Adversarial Signal Editing, a framework for on-the-go forgetting that suppresses the observable influence of private data without modifying model weights. ERASE leverages structured, class-conditioned input perturbations to induce selective forgetting during inference, eliminating the need for retraining, fine-tuning, or model copies. We rigorously characterize sufficient conditions when ERASE provably achieves functional forgetting of designated subclasses while preserving predictions across other subclasses within the same superclass. This analysis offers a principled foundation for inference-time forgetting under mild regularity assumptions. Across diverse architectures and benchmark datasets, ERASE maintains the best observed balance between forgetting efficacy, computational efficiency, and retention fidelity over recent unlearning-based methods. By reimagining data removal as forgetting without unlearning, our work establishes a scalable, regulation-aligned pathway for continual, privacy-conscious learning.

Long-running LLM agents are stateful: beyond the transcript they accrete compressed summaries, plaintext memory, pending tool plans, and, under every serving API, a KV cache. Yet today's "forget" operations delete a plaintext memory record and stop, leaving every artifact derived from the revoked information intact. We formalize execution-state unlearning: after a forget request, the agent must behave as if it had never observed the target. Modeling the runtime as a deterministic transition system, we prove that the pre-target trajectory prefix is shared with this counterfactual world for free, that the post-target suffix is irreducibly tainted without token-level attribution, and that exact unlearning requires at least $T-τ+1$ recomputed transitions, where $τ$ is the target's injection step. Provenance-Guided Selective Replay attains this bound as a cross-layer contract spanning prompt, compressed memory, and cache: a provenance graph locates the injection point, checkpoint restoration reduces to cropping the KV cache, and sanitized replay regenerates the counterfactual suffix. Audited with elicitation, stochastic, and string-free behavioral tests across three agent suites, nine baselines, and three model families, memory deletion leaves leakage unchanged, instruction-based forgetting collapses under elicitation (Leak@probes = 1.00), and source redaction still acts on a revoked preference in 80% of episodes, while selective replay is indistinguishable from a full reset at up to 9x fewer recomputed tokens.

Label removal occurs frequently in classification systems with evolving taxonomies, where categories must be dynamically updated or eliminated. To accommodate such changes, classification models must adapt accordingly. Existing solutions, broadly categorized as retraining-based and feature-space-adjustment-based, share common limitations despite their variations, including reliance on access to original data, substantial computational and storage costs, inconsistent results, poor scalability, and degradation of model utility. To address this, we propose a novel approach that leverages statistical redistribution in the output space to approximate the post-removal confidence vectors of a retrained model. Applicable as a modular output filter, our method bypasses the burden of feature-space adjustments or loss-function convergence, alleviating scalability limitations. Furthermore, by requiring only existing labels and prior output confidences, the method potentially mitigates privacy concerns inherent to data-dependent solutions. Extensive experiments demonstrate competitive performance against full retraining, with improvements in computational efficiency and privacy preservation across several classification tasks.

A long-held intuition in interpretability research is that representational entanglement, the sharing of structure between knowledge domains in a neural network, makes unlearning harder. While the intuition is widespread, it has never been directly tested in a controlled experiment. We present a way to do so: by repurposing Selective Gradient Masking (SGTM), we train a suite of six 254M-parameter language models on English Wikipedia with graded levels of disentanglement between biology and non-biology knowledge. Applying three standard unlearning methods to every model in the suite, we find that more disentangled models consistently achieve better retain-forget trade-offs: at a fixed level of forgetting, the most disentangled models incur roughly $4\times$ lower retain cost under two of the three methods, and $1.3\times$ lower under the third. Because our intervention changes only the model, not the data or the unlearning algorithm, this is direct evidence that representational entanglement is one of the causes of collateral damage in unlearning, as interpretability researchers have long suspected. A similar design could be used to test other structural claims from interpretability.

Class unlearning aims to remove a model's ability to recognize designated forget classes while preserving performance on retain classes. However, low forget accuracy after unlearning does not necessarily mean the class structure has been erased. Approximate unlearning methods can alter classifier decision boundaries while leaving recoverable structure in the representation. Prior work has shown that forget classes can be recovered, but existing approaches require real forget or retain samples, auxiliary data, or reference checkpoints. We study class relearning in a strictly source-free setting, asking whether a forget class can be recovered through a classifier-head update using only the unlearned model. Our approach rests on a theoretical analysis establishing a sufficient alignment condition under which a single gradient step on a synthetic probe set increases the expected logit margin of the forget class. Building on this, we propose a white-box Source-Free Relearning Audit (SFRA), which generates candidate embeddings in representation space and uses model-guided confidence filtering to construct high-confidence retain probes and low-confidence boundary-adjacent probes that are relabelled as the forget class. Gaussian sampling and Softmax confidence are used by default, while ablations with alternative proposal distributions and uncertainty criteria show that recoverability is not specific to these choices. To quantify recoverability, we introduce the Relearning Score (RS), which jointly measures forget-class recovery and retain-accuracy preservation, and report class-matched $Δ$RS relative to a retrained reference. Experiments on CIFAR-10, CIFAR-100, and TinyImageNet with ResNet-18, ViT-B/16, and Swin-T show that several unlearning methods exhibit substantial source-free recoverability, and that for a subset of methods this recoverability exceeds the matched retrained reference.

Machine unlearning for large language models (LLMs) often assumes that a pre-defined forget set matches what the model has memorized, but this frequently breaks in realistic privacy settings where the original training data is inaccessible. We term this gap forget-set misalignment and identify two cases. In Under Unlearning, the forget set omits memorized information and leakage persists. In Out-of-Knowledge Unlearning, the algorithm is driven to "forget" knowledge the model never learned, perturbing parameters and degrading utility. Using gradient-level analysis, we show these behaviors arise from misaligned unlearning targets rather than specific optimization choices. We then propose CONfession-to-Forget-Set (CONFS), a data-blind framework that constructs model-aligned forget sets by eliciting and formalizing the model's memorized knowledge. Across synthetic, multimodal, and real-world benchmarks, CONFS approaches Gold-standard performance on several metrics and achieves a competitive forgetting-utility balance, while preserving utility better than other data-blind forget-set constructions.

Removing a specific individual's information from multimodal large language models (MLLMs) is often needed after deployment, but existing methods rely on a retain set, which is hardest to obtain at that point, and rebuilding it recreates the privacy exposure that unlearning aims to remove. Forgetting from the forget set alone instead damages the shared visual-language computation, harming perception. We cast retain-free unlearning as a localization problem: causal tracing, weight transplant, and Fisher overlap all point to early-to-mid decoder MLPs as the layers where identity information is stored and, unlike other module families, can be modified without substantially disrupting vision. We turn this into Pathway-Aware Visual-attribute Anchoring (PAVA), which confines updates to these layers and pairs a forget loss with a visual-attribute anchor that preserves image-grounded behavior by distilling the model's own pre-unlearning answers from the forget images alone. On MLLMU-Bench and ReMem, PAVA gives the strongest forget-retain trade-off among forget-set-only methods and remains competitive with retain-based baselines.