Fine-Grained Meta Unlearning

• Fine-Grained Meta Unlearning is a framework that targets specific model components—such as hidden activations, gradients, or graph substructures—to erase unwanted knowledge while maintaining performance.
• It employs meta-level signals like mutual information, adversarial gradients, and prototype-based interventions to precisely locate and isolate the memory to be forgotten.
• Empirical studies across LLMs, knowledge graphs, and federated graph learning demonstrate robust deletion effectiveness, minimal utility loss, and resistance to future relearning.

Searching arXiv for papers on fine-grained meta unlearning and closely related unlearning frameworks. Fine-grained meta unlearning denotes a class of machine unlearning methods that seek to remove narrowly specified information while preserving the rest of a model’s functionality, using higher-level signals about where and how forgetting should occur rather than relying only on coarse output-loss manipulation. Across recent work, the term covers several distinct but related settings: representation-guided unlearning in LLMs, meta-learning-based embedding regeneration in knowledge graphs, and fine-grained deletion within federated graph learning. A common theme is selective intervention at the level of hidden activations, gradients, triples, nodes, edges, features, or local substructures, combined with mechanisms intended to preserve retained capability and reduce residual recoverability of the forgotten information (Hu et al., 3 Feb 2025, Xu et al., 2024, Ai et al., 4 Aug 2025).

1. Conceptual scope and problem setting

Fine-grained meta unlearning is motivated by the inadequacy of coarse unlearning procedures that operate only at the output-loss level or through global retraining. In the large-language-model setting, the forget/retain split is formalized through a forget set $D_F$ and a retain set $D_R$, with a standard training-time objective

$$\min_{\theta} \;\mathcal{L}_{MU}(\theta) = \mathbb{E}_{(x,y) \in D_F}\big[\mathcal{L}(y_f \mid x;\theta)\big] + \lambda \,\mathbb{E}_{(x,y)\in D_R}\big[\mathcal{L}(y \mid x;\theta)\big],$$

where $y_f$ is usually a safe or non-informative response and $\lambda$ balances forgetting and retaining (Hu et al., 3 Feb 2025). The difficulty is that such a formulation does not specify where the relevant knowledge is encoded in the network, how forget and retain subspaces overlap, or how to manage conflicting updates.

In knowledge graph embedding unlearning, the objective is similarly selective but structurally localized. A knowledge graph is written as

$$\mathcal{G}=(\mathcal{E},\mathcal{R},\mathcal{T}),\quad \mathcal{T}\subseteq \mathcal{E}\times\mathcal{R}\times\mathcal{E},$$

and the task is to transform embeddings $\mathbf{E}$ into $\mathbf{E}'$ so that performance on the forgetting set $\mathcal{T}_f$ worsens while performance on the remaining set $\mathcal{T}_r$ stays close to the original model and better than retraining only on $D_R$0 (Xu et al., 2024). Here fine granularity refers to triples and their associated local structures rather than whole-entity deletion.

In federated graph learning, fine-grained meta unlearning is defined as structural deletion inside a client’s local subgraph. A meta-unlearn request is

$D_R$1

corresponding to node-level, edge-level, or feature-level deletion inside client $D_R$2’s subgraph $D_R$3 (Ai et al., 4 Aug 2025). The objective is not full client removal, but selective erasure with minimal collateral damage despite message passing and cross-client aggregation.

These settings differ in architecture and data modality, but they share three commitments. First, forgetting is localized to a subset smaller than an entire model or dataset. Second, forgetting is treated as a multi-objective optimization problem with explicit retention constraints. Third, the procedure uses meta-level criteria—information-theoretic measures, task distributions, prototype spaces, or adversarial relearning—to decide how to intervene (Hu et al., 3 Feb 2025, Xu et al., 2024, Ai et al., 4 Aug 2025, Sondej et al., 14 Jun 2025).

2. Information-theoretic and representational guidance in LLMs

A prominent formulation of fine-grained meta unlearning for LLMs is FALCON, a training-time, representation-guided method that moves from coarse loss-level intervention to hidden-activation and gradient-level manipulation (Hu et al., 3 Feb 2025). Its central premise is that forgetting quality depends on identifying layers where forget and retain knowledge are minimally entangled.

For a layer $D_R$4, FALCON defines forget and retain activations $D_R$5 and $D_R$6, and measures their entanglement via mutual information

$D_R$7

For multi-domain forgetting with sub-domains $D_R$8, the aggregate criterion is

$D_R$9

and the selected intervention layer is

$$\min_{\theta} \;\mathcal{L}_{MU}(\theta) = \mathbb{E}_{(x,y) \in D_F}\big[\mathcal{L}(y_f \mid x;\theta)\big] + \lambda \,\mathbb{E}_{(x,y)\in D_R}\big[\mathcal{L}(y \mid x;\theta)\big],$$0

The method estimates mutual information by first reducing activation dimensionality with PCA while keeping 95% variance, then applying Kernel Density Estimation with multivariate Gaussian kernel,

$$\min_{\theta} \;\mathcal{L}_{MU}(\theta) = \mathbb{E}_{(x,y) \in D_F}\big[\mathcal{L}(y_f \mid x;\theta)\big] + \lambda \,\mathbb{E}_{(x,y)\in D_R}\big[\mathcal{L}(y \mid x;\theta)\big],$$1

using Scott’s rule for the bandwidth $$\min_{\theta} \;\mathcal{L}_{MU}(\theta) = \mathbb{E}_{(x,y) \in D_F}\big[\mathcal{L}(y_f \mid x;\theta)\big] + \lambda \,\mathbb{E}_{(x,y)\in D_R}\big[\mathcal{L}(y \mid x;\theta)\big],$$2 (Hu et al., 3 Feb 2025).

This use of mutual information functions as a meta-criterion for parameter selection. It does not merely prescribe what loss to optimize; it prescribes where in the model to optimize it. The paper reports that low-MI layers empirically show less gradient conflict, with cosine similarity between forget and retain gradients closer to zero, and are therefore more suitable for precise unlearning with minimal utility degradation (Hu et al., 3 Feb 2025).

Once the layer is chosen, FALCON performs representation-level separation. At layer $$\min_{\theta} \;\mathcal{L}_{MU}(\theta) = \mathbb{E}_{(x,y) \in D_F}\big[\mathcal{L}(y_f \mid x;\theta)\big] + \lambda \,\mathbb{E}_{(x,y)\in D_R}\big[\mathcal{L}(y \mid x;\theta)\big],$$3, for activations $$\min_{\theta} \;\mathcal{L}_{MU}(\theta) = \mathbb{E}_{(x,y) \in D_F}\big[\mathcal{L}(y_f \mid x;\theta)\big] + \lambda \,\mathbb{E}_{(x,y)\in D_R}\big[\mathcal{L}(y \mid x;\theta)\big],$$4, it computes an SVD,

$$\min_{\theta} \;\mathcal{L}_{MU}(\theta) = \mathbb{E}_{(x,y) \in D_F}\big[\mathcal{L}(y_f \mid x;\theta)\big] + \lambda \,\mathbb{E}_{(x,y)\in D_R}\big[\mathcal{L}(y \mid x;\theta)\big],$$5

and uses the top-$$\min_{\theta} \;\mathcal{L}_{MU}(\theta) = \mathbb{E}_{(x,y) \in D_F}\big[\mathcal{L}(y_f \mid x;\theta)\big] + \lambda \,\mathbb{E}_{(x,y)\in D_R}\big[\mathcal{L}(y \mid x;\theta)\big],$$6 right singular vectors $$\min_{\theta} \;\mathcal{L}_{MU}(\theta) = \mathbb{E}_{(x,y) \in D_F}\big[\mathcal{L}(y_f \mid x;\theta)\big] + \lambda \,\mathbb{E}_{(x,y)\in D_R}\big[\mathcal{L}(y \mid x;\theta)\big],$$7 to construct Principal Offset Vectors,

$$\min_{\theta} \;\mathcal{L}_{MU}(\theta) = \mathbb{E}_{(x,y) \in D_F}\big[\mathcal{L}(y_f \mid x;\theta)\big] + \lambda \,\mathbb{E}_{(x,y)\in D_R}\big[\mathcal{L}(y \mid x;\theta)\big],$$8

These vectors define a direction away from dominant principal subspaces and toward low-variance, less-entangled subspaces (Hu et al., 3 Feb 2025).

For forget data, FALCON uses an InfoNCE-style contrastive loss. If $$\min_{\theta} \;\mathcal{L}_{MU}(\theta) = \mathbb{E}_{(x,y) \in D_F}\big[\mathcal{L}(y_f \mid x;\theta)\big] + \lambda \,\mathbb{E}_{(x,y)\in D_R}\big[\mathcal{L}(y \mid x;\theta)\big],$$9 is the updated-model activation, $y_f$0 the POV target, and $y_f$1 negatives from the frozen pre-unlearning model, the similarities are

$y_f$2

and the forget loss is

$y_f$3

For retain data, it uses cosine alignment between updated and frozen activations,

$y_f$4

The intended effect is explicit: forget representations are pushed away from the old harmful subspace and toward a POV-defined subspace, while retain representations remain aligned with the frozen model (Hu et al., 3 Feb 2025).

This representation-level formulation is fine-grained in a stronger sense than simple token-logit suppression. It attempts to reshape hidden geometry so that harmful knowledge is no longer supported by the internal representation, a claim reinforced in that work by Logit Lens probing and resistance to Enhanced GCG recovery attempts (Hu et al., 3 Feb 2025).

3. Gradient-level meta control and robustness to relearning

A second major line of fine-grained meta unlearning in LLMs emphasizes robustness against post-hoc relearning. MUDMAN treats unlearning as a process that must remain effective even after an attacker fine-tunes the model again on the forget set (Sondej et al., 14 Jun 2025). Its key claim is that many conventional unlearning procedures produce dormant rather than erased circuits, allowing rapid recovery under later fine-tuning.

MUDMAN maintains a main model with parameters $y_f$5 and an adversary model with parameters $y_f$6. The main model is trained on retain data, while the adversary is repeatedly trained on forget data to relearn the unwanted capability. This adversary supplies a meta-unlearning signal approximating the gradients an attacker would later exploit. With retain mini-batches $y_f$7 and forget mini-batches $y_f$8, the retain gradient is

$y_f$9

and a momentum accumulator is maintained: $\lambda$0 The adversary is updated on forget data by

$\lambda$1

and the unlearning gradient is then computed from an unlearning loss such as negative cross-entropy or negative entropy: $\lambda$2 This is the meta-learning component: the model is not only unlearning current behavior, but using an adversary to approximate future relearning trajectories (Sondej et al., 14 Jun 2025).

MUDMAN adds two fine-grained controls. The first is global L2 normalization of the unlearning gradient,

$\lambda$3

which the paper identifies as necessary because raw unlearning gradients can shrink substantially late in training (Sondej et al., 14 Jun 2025). The second is Disruption Masking, which retains only parameter updates whose sign agrees with the momentum-smoothed retain gradient: $\lambda$4 The update is then

$\lambda$5

This fine-grained gating operates at the level of individual weight components. Its purpose is to avoid “damage and repair” dynamics by allowing only non-disruptive updates—those that align with longer-term retain gradients (Sondej et al., 14 Jun 2025). The method is typically applied only to selected first-MLP modules such as gate projections or up projections, reflecting the paper’s mechanistic intuition that early MLP layers determine which neurons are active and can therefore make relevant pathways harder to resurrect under relearning (Sondej et al., 14 Jun 2025).

The reported evaluation protocol is explicitly robustness-oriented. Unlearning on CodeSearchNet or Pile-Bio is followed by a relearning stage on the forget data, and success is measured by the model’s continued poor performance on forget tasks after that relearning. On WMDP-Bio, the paper states that MUDMAN outperforms TAR by 40% under this robust unlearning metric, while preserving retain performance under tight constraints on LM loss or MMLU drop (Sondej et al., 14 Jun 2025). This suggests a narrower interpretation of fine-grained meta unlearning: not only selective erasure, but selective modification of weight space so that future recovery gradients become less effective.

4. Meta-learning-based fine-grained unlearning in knowledge graphs

In knowledge graph embedding, fine-grained meta unlearning takes a different form. MetaEU frames unlearning as a bi-level or meta-learning problem over many synthetic unlearning tasks, allowing the model to adapt to new forget requests without retraining from scratch (Xu et al., 2024). The high-level meta-objective is described conceptually as MAML-like: $\lambda$6 where $\lambda$7 is obtained by adaptation on the support set of task $\lambda$8 (Xu et al., 2024).

The learned meta-knowledge is not tied to particular entities or triples. Rather, it captures relation-pattern information and neighborhood structure patterns through two modules. The Relation-Aware Entity Embedding Generator (RAEEG) constructs an initial entity embedding from incoming and outgoing relation embeddings: $\lambda$9 The Neighbor-Enhanced Embedding Modulator (NEEM) then refines this via a relational GNN,

$$\mathcal{G}=(\mathcal{E},\mathcal{R},\mathcal{T}),\quad \mathcal{T}\subseteq \mathcal{E}\times\mathcal{R}\times\mathcal{E},$$0

and a hierarchical embedding integrator aggregates multi-layer representations: $$\mathcal{G}=(\mathcal{E},\mathcal{R},\mathcal{T}),\quad \mathcal{T}\subseteq \mathcal{E}\times\mathcal{R}\times\mathcal{E},$$1 Together, RAEEG and NEEM define a structure-aware embedding generator that can regenerate or perturb entity embeddings on the basis of local context, including for unseen entities (Xu et al., 2024).

MetaEU’s objective combines retention and forgetting at the ensemble level. With base learners $$\mathcal{G}=(\mathcal{E},\mathcal{R},\mathcal{T}),\quad \mathcal{T}\subseteq \mathcal{E}\times\mathcal{R}\times\mathcal{E},$$2 and weights $$\mathcal{G}=(\mathcal{E},\mathcal{R},\mathcal{T}),\quad \mathcal{T}\subseteq \mathcal{E}\times\mathcal{R}\times\mathcal{E},$$3, the retention-oriented ensemble learning loss is

$$\mathcal{G}=(\mathcal{E},\mathcal{R},\mathcal{T}),\quad \mathcal{T}\subseteq \mathcal{E}\times\mathcal{R}\times\mathcal{E},$$4

while the forgetting-oriented ensemble unlearning loss is

$$\mathcal{G}=(\mathcal{E},\mathcal{R},\mathcal{T}),\quad \mathcal{T}\subseteq \mathcal{E}\times\mathcal{R}\times\mathcal{E},$$5

These are combined as

$$\mathcal{G}=(\mathcal{E},\mathcal{R},\mathcal{T}),\quad \mathcal{T}\subseteq \mathcal{E}\times\mathcal{R}\times\mathcal{E},$$6

with additional terms $$\mathcal{G}=(\mathcal{E},\mathcal{R},\mathcal{T}),\quad \mathcal{T}\subseteq \mathcal{E}\times\mathcal{R}\times\mathcal{E},$$7 and $$\mathcal{G}=(\mathcal{E},\mathcal{R},\mathcal{T}),\quad \mathcal{T}\subseteq \mathcal{E}\times\mathcal{R}\times\mathcal{E},$$8 for regularization of forgetting strength and task-specific adaptation (Xu et al., 2024).

In this formulation, the granularity of forgetting lies in triples and their associated local structures. The model does not simply delete all appearances of an entity or globally corrupt an embedding. Instead, it re-generates entity embeddings in a way that preserves behavior on triples in $$\mathcal{G}=(\mathcal{E},\mathcal{R},\mathcal{T}),\quad \mathcal{T}\subseteq \mathcal{E}\times\mathcal{R}\times\mathcal{E},$$9 while degrading behavior on patterns associated with $\mathbf{E}$0 (Xu et al., 2024). This is a meta-unlearning framework because the system is trained across many subgraph tasks to learn reusable strategies for how to unlearn future subsets.

The reported experiments use FB15k-237 with $\mathbf{E}$1 task subgraphs, 10,000 for meta-training and 200 for validation, and evaluate Test and Forget performance with MRR and Hits@$\mathbf{E}$2 averaged over 10 runs (Xu et al., 2024). For TransE, the paper reports RAW Test MRR 0.7254, Retrained Test MRR 0.6918, and Unlearned Test MRR 0.7153, while Forget MRR drops from 0.7157 in RAW and 0.2473 in Retrained to 0.1740 in Unlearned (Xu et al., 2024). The stated interpretation is that meta-guided unlearning more effectively removes the forgotten triples while preserving global predictive performance better than retraining.

5. Federated graph learning and intra-client meta unlearning

Federated graph unlearning extends the notion of fine-grained meta unlearning to decentralized graph systems. In PAGE, the meta-unlearning component handles within-client deletion requests for specific nodes, edges, or feature subsets, while preserving the remainder of the client’s data and the utility of other clients (Ai et al., 4 Aug 2025). This is distinct from complete client unlearning.

The federated training objective is written as

$\mathbf{E}$3

with standard FedAvg over clients $\mathbf{E}$4 (Ai et al., 4 Aug 2025). For meta unlearning, PAGE uses a three-stage pipeline: Prototype Matching for Local Unlearn, Adversarial Graph Generation, and Negative Knowledge Distillation for influenced clients.

At each client $\mathbf{E}$5, class prototypes are constructed as

$\mathbf{E}$6

For the deletion request at client $\mathbf{E}$7, a prototype of the to-be-deleted data is formed: $\mathbf{E}$8 The server builds a global prototype basis from non-unlearning clients using Gram–Schmidt orthogonalization, collects the basis vectors into $\mathbf{E}$9, and projects the deletion prototype: $\mathbf{E}'$0 It then defines the prototype gradient

$\mathbf{E}'$1

which is orthogonal to the shared prototype basis: $\mathbf{E}'$2 The requesting client then performs local unlearning by minimizing

$\mathbf{E}'$3

The stated purpose is to move the representation of deleted data toward the orthogonal private direction, thereby scrambling the latent representation of the forgotten subset while respecting globally shared structure (Ai et al., 4 Aug 2025).

Local forgetting alone is treated as insufficient. PAGE next generates adversarial graphs that maximize the difference between the original local model $\mathbf{E}'$4 and the locally unlearned model $\mathbf{E}'$5. With continuous feature and adjacency variables, the discrepancy loss is

$\mathbf{E}'$6

the regularizer is

$\mathbf{E}'$7

and the adversarial objective is

$\mathbf{E}'$8

with $\mathbf{E}'$9, $\mathcal{T}_f$0, and top-$\mathcal{T}_f$1 edge perturbations with $\mathcal{T}_f$2 used to construct the final adversarial graph $\mathcal{T}_f$3 (Ai et al., 4 Aug 2025). This graph acts as a probe for residual knowledge.

For other potentially influenced clients, PAGE then applies negative knowledge distillation. The positive preservation loss is

$\mathcal{T}_f$4

the negative unlearning loss is

$\mathcal{T}_f$5

and the final objective is

$\mathcal{T}_f$6

This is intended to preserve local utility on the client’s own data while aligning its behavior on sensitive adversarial probes with the unlearned reference model (Ai et al., 4 Aug 2025).

The evaluation covers node classification on Cora, PubMed, Coauthor CS, Amazon Photo, Tolokers, Minesweeper, Amazon-ratings, ogbn-arxiv, Coauthor Physics, and Amazon Computers, with 10% node-, edge-, or feature-level meta-unlearning and accuracy measured on local data (Ai et al., 4 Aug 2025). PAGE is reported to improve prediction accuracy by about 1.50% on average over state-of-the-art meta-unlearning methods, and under poisoned edge attacks on Photo with GAT and SGC, it yields 2.94%–8.62% higher accuracy than competing graph unlearning methods (Ai et al., 4 Aug 2025). The same paper further reports that its influence module can be added to existing GU methods, improving prediction accuracy by 2.82% on Cora and 4.49% on CS, and improving unlearning performance by 4.64% and 7.22%, respectively (Ai et al., 4 Aug 2025).

6. Evaluation regimes, utility preservation, and resistance to recovery

Across these frameworks, evaluation is not limited to whether the target examples become harder. Fine-grained meta unlearning is typically assessed along at least three dimensions: deletion effectiveness, retained-task utility, and robustness against recovery or residual leakage.

In FALCON, deletion effectiveness is measured primarily with WMDP on malicious-use domains, utility with MMLU and WikiText perplexity, and robustness with Logit Lens probing and Enhanced GCG attacks (Hu et al., 3 Feb 2025). On Zephyr-7B-Beta, base performance is reported as WMDP-Bio 63.7, WMDP-Cyber 43.8, MMLU 58.1, and PPL 1.5, while FALCON reaches Bio 26.7, Cyber 25.3, MMLU 57.4, and PPL 1.5. On Yi-6B-Chat, base performance is Bio 65.4, Cyber 42.6, MMLU 61.8, PPL 1.5, and FALCON reaches Bio 27.7, Cyber 25.3, MMLU 60.3, PPL 1.5. On Mistral-7B-Instruct-v0.3, the corresponding values are 66.9, 41.9, 59.7, 1.4 for the base model and 28.0, 24.3, 57.9, 1.4 after unlearning (Hu et al., 3 Feb 2025). The paper’s claim is that strong WMDP forgetting is achieved with MMLU and perplexity close to the base model using single-layer manipulation.

The recovery results are central to its argument. Under Enhanced GCG on Yi-6B-Chat, starting from unlearned scores Bio 27.7 and Cyber 25.3, attack iterations from 500 to 2000 produce Bio approximately 27.6–28.9 and Cyber approximately 24.7–26.4, which the authors interpret as no significant recovery (Hu et al., 3 Feb 2025). This supports the paper’s contention that the forgetting is representation-level rather than merely behavioral.

MUDMAN uses a stricter post-unlearning relearning protocol. Unlearning is followed by fixed-budget supervised fine-tuning on the forget set, and success requires that forget loss or WMDP-Bio degradation remains high after that attack (Sondej et al., 14 Jun 2025). The retain side is constrained by small allowable increases in retain LM loss and at most a 1 percentage point drop in MMLU accuracy in WMDP experiments (Sondej et al., 14 Jun 2025). The reported conclusion is that meta-learning, Disruption Masking, and normalization are jointly required for robust irreversibility, with Disruption Masking making the largest contribution (Sondej et al., 14 Jun 2025).

In MetaEU, utility preservation is quantified by Test MRR and Hits@$\mathcal{T}_f$7 on remaining triples, while forgetting is quantified by degraded performance on the Forget set (Xu et al., 2024). The paper emphasizes that Unlearned embeddings remain close to RAW on Test performance and outperform Retrained, while exhibiting the lowest scores on the Forget set across TransE, DistMult, ComplEx, and RotatE (Xu et al., 2024). This evaluation regime encodes the intended fine granularity: forgetting should be stronger than retraining-only baselines on the target subset while utility loss on unrelated structure remains lower.

PAGE evaluates local model accuracy after node-, edge-, and feature-level deletion and uses membership inference attack as a post-hoc verification of meta-unlearning quality (Ai et al., 4 Aug 2025). The absence of certified guarantees is acknowledged explicitly; the evidence is empirical, based on retrain proximity, MIA behavior, and robustness under attack (Ai et al., 4 Aug 2025).

7. Relation to adjacent unlearning paradigms, limitations, and open questions

Fine-grained meta unlearning should be distinguished from fine-grained unlearning without meta-learning. EFUF, for instance, is an efficient fine-grained unlearning framework for object-hallucination mitigation in multimodal LLMs, but it is explicitly not a meta-learning or meta-unlearning method (Xing et al., 2024). It performs object-level and subsentence-level selection using CLIP-based scoring and applies gradient ascent on hallucinated fragments together with supervised descent on grounded fragments and sentence-level examples. Its objective is

$\mathcal{T}_f$8

with $\mathcal{T}_f$9 and $\mathcal{T}_r$0, and only mapping layers are updated for 1 epoch (Xing et al., 2024). The paper reports consistent reductions in CHAIR and human hallucination metrics across MiniGPT4, mPLUG-owl, LLaVA, and ShareGPT4V, while preserving or improving BLEU, informativeness, and fluency (Xing et al., 2024). This suggests that fine granularity alone does not imply meta unlearning; the “meta” component requires a higher-order mechanism for selecting interventions or generalizing forgetting strategies.

A broader comparison highlights several recurring fault lines.

Framework	Fine-grained unit	Meta mechanism
FALCON (Hu et al., 3 Feb 2025)	Hidden activations, gradients, selected layer	MI-based layer selection and orthogonal gradient unalignment
MUDMAN (Sondej et al., 14 Jun 2025)	Weight components in selected modules	Adversary-driven meta-unlearning against future relearning
MetaEU (Xu et al., 2024)	Triples and local KG structures	Task-distributed meta-learning with RAEEG and NEEM
PAGE (Ai et al., 4 Aug 2025)	Nodes, edges, features within a client	Prototype-space isolation and influence-aware distillation
EFUF (Xing et al., 2024)	Object mentions and local subsentences	No meta-learning

Several limitations recur across the literature. FALCON is evaluated only on text LLMs at 7B scale, uses approximate MI estimation through PCA and KDE, and operates at a single selected layer in reported experiments (Hu et al., 3 Feb 2025). MUDMAN’s guarantees are empirical and tied to a particular fixed-budget relearning protocol; its success depends on carefully selected modules and hyperparameter search with Optuna over hundreds of trials (Sondej et al., 14 Jun 2025). MetaEU is evaluated only on FB15k-237, introduces meta-learning and ensemble complexity, and does not fully specify exact inner- versus outer-loop optimization details (Xu et al., 2024). PAGE assumes that prototypes adequately summarize private versus shared knowledge and that Gram–Schmidt decomposition in prototype space is meaningful, while providing no formal convergence, differential privacy, or certified unlearning guarantees (Ai et al., 4 Aug 2025).

A common misconception is that selective forgetting can be adequately achieved by maximizing loss on forget data while minimizing loss on retain data. The surveyed work collectively argues otherwise, though in different ways. FALCON contends that output-level loss balancing is too coarse because it ignores representation entanglement and gradient conflict (Hu et al., 3 Feb 2025). MUDMAN argues that such methods often produce dormant circuits that are easily reactivated (Sondej et al., 14 Jun 2025). MetaEU shows that retraining on remaining data can still leave enough structural signal to reconstruct forgotten triples (Xu et al., 2024). PAGE argues that in federated graph learning, local forgetting is insufficient because deleted knowledge may already have permeated into other clients (Ai et al., 4 Aug 2025).

Taken together, these works suggest that fine-grained meta unlearning is less a single algorithmic family than a design principle: identify the smallest operative unit of unwanted knowledge, introduce a meta-level signal that localizes or anticipates forgetting, and enforce retention not only through scalar loss balancing but through representational, geometric, or cross-client constraints. A plausible implication is that future systems will combine information-theoretic layer selection, adversarial recovery models, structure-aware regeneration, and influence tracing across distributed systems rather than treating unlearning as a single-step optimization problem.