In this paper
AbstractIntroductionRelated WorksMemento: Prompt-Space Recovery of Edited KnowledgePrompt-Space RecoveryContinuous Recovery Training ObjectiveDiscrete Recovery Training ObjectiveExperimentsEvaluation MetricsEditing EfficacyPrefix Length SelectionRecovery EffectivenessText-Only RecoveryFactual Recall Beyond Model EditingMechanistic AnalysisConclusionReproducibilityModelsEditing TechniquesContinuous RecoveryDiscrete RecoveryComputeCost AnalysisComplexity AnalysisAblations on the Recovery ObjectiveLLM-as-a-Judge EvaluationLLM-as-a-Judge System PromptReferencesWorksComputer Science
The Truth Beneath the Tokens: Prompt-Space Recovery of Knowledge Edits in Language Models
Abstract
Model editing provides an efficient way to update factual knowledge in transformer models. Recent approaches reverse edits by modifying, masking, or decomposing model weights, but make recovery dependent on weight-space interventions. We instead show that pre-edit knowledge can be recovered through input-space interventions alone, without modifying the edited model's weights. We introduce Memento, a prompt-space mechanism for auditing residual knowledge that learns a reusable recovery signal to steer an edited model toward its pre-edit response. Memento supports both continuous recovery prefixes and discrete textual prompts for standard text interfaces. Across three LLMs and five editing techniques, Memento recovers 80.07% of held-out edited facts, achieving a 9.58% relative improvement over baselines, suggesting that editing may suppress the facts without fully eliminating their recoverability. We further investigate whether prompt-space recovery extends beyond model editing. On pristine, unedited LLMs, Memento improves average factual recall by 53.9% relative to standard inference. These results demonstrate the potential of prompt-space recovery for auditing residual knowledge in edited models and investigating the gap between factual knowledge and its accessibility during inference. We will release our source code.
Introduction
Model editing is used for correcting factual content in large language models (LLMs) after training, including removal of personally identifiable information [21, 22], and correction of outdated or incorrect facts [18, 19]. Retraining from scratch offers the strongest guarantee but is infeasible at deployment scale [12]; finetuning is more practical but remains expensive for large models [8]; and retrieval-augmented generation moves knowledge control outside the model, introducing failure modes of its own [1, 6]. Model editing instead performs precise, localized parameter updates, typically following a locate-then-edit paradigm, where it identifies the weights responsible for a target association, and then rewrites them.
The success of an edit is typically evaluated behaviorally, that is, if the model produces the new target when prompted, the edit is considered successful [18, 27]. This has also motivated a line of work on edit reversal. One Mask [11] learns a shared sparse binary mask over edited MLP weights to reverse edits across multiple facts. Youssef et al. [24] show that edited information is concentrated in the dominant singular subspace of the edited weight matrix. Jiang et al. [13] study the selective reversal by learning element-wise gates over singular vectors to reverse targeted facts while preserving other edits. These approaches, however, formulate recovery as weight-space reversal, modifying model weights to undo the effects of an edit. Consequently, they do not establish whether pre-edit knowledge remains recoverable from the edited model without reversing the parameter update. This leaves a fundamental question unexplored: does editing truly eliminate the pre-edit knowledge, or does it primarily suppress its accessibility under standard prompting?
We introduce MEMENTO, a novel prompt-space auditing mechanism for recovering pre-edit knowledge from edited LLMs. Rather than modifying model parameters, MEMENTO learns a recovery signal that steers an edited model toward its pre-edit factual response. MEMENTO can be trained in both continuous and discrete prompt spaces. To elicit the pre-edit response, the continuous variant learns a soft input representation, while the discrete variant learns a textual recovery prompt that is prepended to the input. Across three models and five editing techniques, continuous MEMENTO recovers 80.07% of held-out edited facts, achieving a 9.58% relative improvement over One Mask Holmov et al., [11]. Our discrete variant achieves 49.0% recovery using a textual recovery prompt, making it compatible with standard API interfaces once the prompt has been learned. Together, these results suggest that editing may suppress factual knowledge without fully eliminating its recoverability.
Finally, we investigate whether MEMENTO can surface factual knowledge beyond editing. Recent work Calderon et al., [5] distinguishes factual knowledge encoded in model weights from what is accessible through standard inference. Building on this distinction, we evaluate whether a learned prompt-space recovery signal can improve factual recall in pristine, unedited models. Across three LLMs, MEMENTO improves average recall by 53.9% relative to standard inference. While prior work characterizes the gap between factual encoding and recall, our results show that a reusable prompt-space intervention elicits facts that standard prompting fails to retrieve.
Our contributions are summarized as follows:
We introduce MEMENTO, a prompt-space auditing mechanism to recover pre-edit knowledge in LLMs without modifying their parameters. MEMENTO learns reusable recovery signals in both continuous and discrete prompt spaces, including a textual recovery prompt that can be supplied through standard language-model interfaces.
We evaluate MEMENTO on multiple LLMs and editing techniques. The continuous variant recovers 80.07% of the held-out facts, a 9.58% relative improvement over the baseline, while the discrete variant achieves 49.0% recovery, showing that pre-edit knowledge can be recovered even via textual signal.
We show that MEMENTO can also improve factual recall in pristine, unedited LLMs. Across three models, it improves average recall by 53.9% relative to standard inference, suggesting that prompt-space recovery can help probe the gap between factual knowledge a model may contain and what it retrieves under standard prompting.
Related Works
Model Editing. Model editing aims to update factual associations in LLMs without retraining from scratch. Editing techniques divide into parameter-preserving approaches, which add memory modules or rely on in-context abilities [20, 10, 26], and parameter-modifying techniques, like ROME and MEMIT [18, 19], which modify the weights. Subsequent approaches improve edit locality and preservation through null-space projection, weight regularization, and constrained matrix updates [8, 9, 17]. However, works on ripple effects and multi-hop reasoning further show editing can have unintended consequences beyond targeted facts [7, 4 a]. We study the parameter-modifying family, since an edit that never enters the weights raises no question about what the weights retain.
Edit Reversal and Recovery. Recent work investigates if model edits can be reversed to restore pre-edit behavior. One Mask Holmov et al., [11] learns a shared sparse mask over edited weights to reverse edits across multiple facts. Youssef et al. [24] propose a reversal method based on bottom-rank approximations of edited weight matrices. Jiang et al. [13] introduce gated singular-vector shrinkage for selectively reversing edits. These methods suggest that a lot of pre-edit behavior can be recovered via weight-space interventions. Prompt-based reversal has been explored for in-context edits [23]. Baser et al. [3] recover pre-edit responses by searching over alternative tokenizations of a query without modifying the weights. However, its reconstruction procedure incurs additional inference overhead, requiring multiple generations over alternative tokenizations for each target fact.
Prompt-Based Optimization. Prompt-based optimization steers frozen LLMs through optimized input representations. Continuous methods, such as prompt tuning and prefix tuning, learn trainable input embeddings while keeping the weights fixed [16, 15]. Discrete methods optimize sequences of vocabulary tokens. For example, Zou et al. [28] use gradient-guided search to construct adversarial suffixes that elicit targeted behaviors. Yu et al. [25] further show that optimizing a single continuous prefix emembedding can refine model behavior while keeping the backbone frozen. MEMENTO adopts prompt-based optimization for targeting recovery of pre-edit knowledge. It learns a continuous recovery prefix or searches for a discrete textual signal, evaluating whether the signal generalizes to different editing techniques.
Memento: Prompt-Space Recovery of Edited Knowledge
MEMENTO is a prompt-space approach to recovering pre-edit factual behavior from LLMs without modifying their parameters. As illustrated in Figure 1, it learns a reusable recovery signal in continuous or discrete prompt space, optimized to restore original factual associations.
Figure 1. Overview of MEMENTO’s continuous prompt-space recovery. A trainable prefix is prepended to factual prompts and optimized across training facts to recover the original object from an edited model , whose parameters remain frozen. The objective combines supervised recovery (), preference optimization (), fluency preservation (), and distillation from the pre-edit model (). Gradients update only ; the pre-edit model supplies distillation targets during training and is not needed at inference. Without the prefix, the edited LLM answers the edited object (); with it, the LLM produces the original object ().
We represent a factual association as a triple , where denotes a subject, a relation, and the original object, and denotes the corresponding factual prompt. Given a model and an editing request , where denotes the target object that replaces , a model editing technique produces an edited model
such that the edited model preferentially predicts in response to . Our goal is to recover the original object from the edited model without modifying its weights. We refer to this as knowledge recovery, distinguishing it from weight-space edit reversal, which intervenes directly on the model weights to counteract the effects of an edit. We analyze training and per-query costs, together with their asymptotic complexity, in Appendix § 7.
Prompt-Space Recovery
We hypothesize that editing can suppress the retrieval of a factual association without completely eliminating its accessibility. If so, an appropriate input intervention may be able to elicit the original association. MEMENTO learns a continuous recovery prefix , where is the prefix length and is the model’s embedding dimension. Given the input embedding matrix and tokenized factual prompt , the resulting input is
where denotes concatenation. A single prefix shared across a collection of edited facts and optimized with an objective function as
This prefix captures a shared recovery direction rather than storing the answer to individual edits.
Continuous Recovery Training Objective
We train the continuous recovery prefix using four complementary loss terms, each serving a distinct purpose: (i) Recovery loss () increases the likelihood of the pre-edit response; (ii) Preference loss () explicitly encourages the pre-edit response to be preferred over the edited target; (iii) Fluency loss () limits changes to the model’s behavior on inputs unrelated to the edited facts; and (iv) Distillation loss () aligns the edited model’s predictions under the recovery prefix with those of the pre-edit model. Together, these objectives encourage factual recovery while constraining unintended changes to model behavior.
For a prefix and a continuation , we define the length-normalized log-likelihood under as
where for continuous recovery. Let and denote the token sequences obtained by tokenizing the original object and edited target , respectively. increases the likelihood of :
However, increasing its likelihood alone does not explicitly encourage the original object to be preferred over the edited target. Hence, we introduce to separate their likelihoods:
where denotes the sigmoid function, controls the preference scale, and specifies the desired margin. explicitly encourages the original object to receive a higher likelihood than the edited object, reducing the chance of the recovery prefix favoring the post-edit association.
We further introduce a fluency regularizer:
where denotes generic text unrelated to the edited facts, and denotes the Kullback-Leibler divergence from one probability distribution to another. penalizes changes in output distribution caused by adding the recovery prefix to unrelated inputs. It thus discourages the prefix from altering behavior beyond the factual prompts for which recovery is intended.
We use to align the edited model under the recovery prefix with the pre-edit model:
where denotes the next-token probability distribution produced by the pre-edit model, with its logits scaled by the temperature . is defined analogously for the edited model with the recovery prefix. The KL divergence is computed over the vocabulary at the final prompt position, and controls the softness of the distributions. The overall objective is
The coefficients , , and control the relative contributions of the losses. We present their individual effects in Appendix C.
Discrete Recovery Training Objective
Let be the model’s token vocabulary. The continuous prefix is represented in the model’s embedding space. To obtain a textual recovery prompt, Memento learns a token sequence
which is prepended to the query. The prefix is optimized over the same collection of edited facts:
where in likelihood terms. requires a single prefix to prefer original object over edited object across multiple facts. Let be the tokenizer’s decoding function, and be the corresponding tokenization function. The search is constrained to text-realizable token sequences:
so that the learned sequence is preserved when decoded and re-tokenized by a text interface.
We optimize Eq. (11) over using greedy coordinate gradient (Zou et al., [28]). The search iteratively proposes token substitutions ranked by the gradient with respect to one-hot token indicators, and evaluates feasible candidates exactly on batches of edited facts. The resulting sequence is the one with the lowest on a subset of withheld from the search. Hence, a single learned prefix can be used to elicit the pre-edit responses through a standard textual interface.
Experiments
Training Details. We follow the setup of One Mask (Holmov et al., [11]) and draw facts from the training and test splits of CounterFact ([18]). Following the configuration specified in the official One Mask repository, we use 1,000 facts for training and 1,000 held-out facts for evaluation. We evaluate five model editing techniques: MEMIT ([19]), AlphaEdit ([8]), PRUNE ([17]), RECT ([9]), and CoME ([14]).
We evaluate MEMENTO on GPT2-XL (1.5B), GPT-J (6B), and Llama3 (8B, Instruct), and compare its recovery performance with One Mask. For the editing techniques, we first edit the model using the corresponding training set, and train the recovery prefix on the resulting edited model. The learned prefix is then evaluated on held-out facts from the test set (excluded from training).
We consider two evaluation settings. In the within-method setting, the recovery prefix is trained and evaluated on models edited using the same editing technique. In the cross-method setting, the prefix is trained using one technique and evaluated on a model edited using another. Specifically, we train recovery prefixes on each editing technique and evaluate them across other techniques. This tests whether the recovery signal captures properties shared across editing techniques rather than artifacts specific to a particular one. All hyperparameters and compute details for reproducibility are given in the Appendix § 6.
Evaluation Metrics
Top-1 Recovery (Top-1). Our primary metric measures whether MEMENTO restores the pre-edit model’s top prediction, following One Mask ([11]). For each factual prompt, we compare the top prediction of the edited model with the recovery prefix against that of the corresponding pre-edit model. We report the percentage of held-out facts for which the predictions match. Higher scores indicate more successful behavioral recovery. We use top-1 in both the within-method and cross-method settings, referring to the latter as Cross-Method Recovery (CMR).
LLM-as-a-Judge Evaluation. For our research question 3, we evaluate factual recall on pristine, unedited models using an LLM judge. Given the factual prompt, its reference answer, and the model’s generated continuation, the judge determines whether the continuation correctly expresses the requested fact. This semantic evaluation accommodates equivalent entity names and answers expressed indirectly, rather
than requiring an exact lexical match. We use deepseek-ai/DeepSeek-V4.1-Flash as the judge and report the percentage of continuations judged correct. The model’s generation uses FP32 inference and a maximum of 32 output tokens. Appendix § 9 provides the judging instructions and illustrative examples.
Canonical-Answer Correctness. For Research Question 3, we complement semantic judging with a stricter, rule-based metric. A continuation is counted as correct only if it begins with the canonical reference answer after Unicode, case, punctuation, and whitespace normalization. We report the percentage of continuations satisfying this criterion. This metric is deterministic and easy to reproduce, but may classify semantically valid answers as incorrect when they use aliases or express the answer indirectly.
Editing Efficacy
Before evaluating recovery, we verify that the editing techniques successfully modify the models’ factual predictions. Table 1 reports editing efficacy on the held-out facts across all three models and five editing techniques. All techniques achieve high efficacy, with scores ranging from 0.85 to 0.97. These results show that the evaluated models have been successfully edited, providing a meaningful basis for assessing whether MEMENTO can recover pre-edit behavior.
| Technique | GPT2-XL | GPT-J | Llama3 |
| AlphaEdit | 0.97 | 0.97 | 0.92 |
| MEMIT | 0.95 | 0.97 | 0.91 |
| RECT | 0.90 | 0.95 | 0.85 |
| PRUNE | 0.96 | 0.91 | 0.96 |
| CoME | 0.95 | 0.94 | 0.90 |
Table 1. Efficacy of all editing techniques across the models being evaluated over the held-out facts.
Prefix Length Selection
The recovery prefix length determines the number of trainable input embeddings available to steer the edited model toward its pre-edit behavior. We select the prefix length using a calibration experiment with MEMIT-edited models. We evaluate on all the three models, measuring top-1 recovery on each. As shown in Table 2, increasing the prefix length from 1 to 2 substantially improves recovery across all three models, indicating that a single input embedding provides limited capacity to recover the original factual associations. Performance generally improves with longer prefixes, although the gains are not monotonic and vary across models. In particular, achieves top-1 recovery scores of 0.79, 0.80, and 0.84 on GPT2-XL, GPT-J, and Llama3, respectively. While achieves a marginally higher score on Llama3 (0.85), provides consistently strong performance across all three architectures. Further increasing the prefix length to 32 or 64 does not yield consistent improvements, with the 64-token prefix notably reducing recovery on Llama3. Thus, we select as a common prefix length for all subsequent experiments, without introducing model-specific configurations.
| GPT2-XL | GPT-J | Llama3 | |
| 1 | 0.58 | 0.54 | 0.57 |
| 2 | 0.73 | 0.71 | 0.70 |
| 4 | 0.71 | 0.68 | 0.77 |
| 8 | 0.75 | 0.77 | 0.85 |
| 16 | 0.79 | 0.80 | 0.84 |
| 32 | 0.76 | 0.77 | 0.82 |
| 64 | 0.77 | 0.72 | 0.58 |
Table 2. Top-1 Recovery for different prefix lengths on MEMIT-edited models.
Recovery Effectiveness
Research Question 1
Can prompt-space interventions recover pre-edit behavior from knowledge-edited LLMs without modifying their weights?
Table 3 presents the results for within-method top-1 and cross-method recovery (CMR).
| Model | Method | MEMIT Top-1 | MEMIT CMR | AlphaEdit Top-1 | AlphaEdit CMR | PRUNE Top-1 | PRUNE CMR | CoME Top-1 | CoME CMR | RECT Top-1 | RECT CMR |
| GPT2-XL | One Mask | 0.79 | 0.76 | 0.78 | 0.68 | 0.77 | 0.76 | 0.83 | 0.78 | 0.79 | 0.79 |
| MEMENTO | 0.79 | 0.77 | 0.80 | 0.78 | 0.72 | 0.74 | 0.77 | 0.75 | 0.80 | 0.78 | |
| GPT-J | One Mask | 0.72 | 0.72 | 0.71 | 0.70 | 0.73 | 0.72 | 0.72 | 0.70 | 0.78 | 0.75 |
| MEMENTO | 0.80 | 0.78 | 0.82 | 0.77 | 0.74 | 0.77 | 0.79 | 0.78 | 0.81 | 0.82 | |
| Llama3 | One Mask | 0.62 | 0.56 | 0.78 | 0.68 | 0.64 | 0.57 | 0.65 | 0.59 | 0.65 | 0.69 |
| MEMENTO | 0.84 | 0.82 | 0.85 | 0.83 | 0.78 | 0.78 | 0.84 | 0.80 | 0.86 | 0.81 |
Table 3. Top-1 recovery and CMR of MEMENTO and One Mask across three LLMs and five editing techniques.
Within-Method Recovery. MEMENTO achieves strong recovery performance across the evaluated models, with particularly consistent improvements on GPT-J and Llama3, outperforming One Mask across all five editing techniques. The largest gain occurs with MEMIT on Llama3, where recovery increases from 0.62 to 0.84. The results on GPT2-XL are more mixed. MEMENTO improves top-1 recovery for AlphaEdit (0.78 to 0.80) and RECT (0.79 to 0.80), matches One Mask on MEMIT (0.79), and underperforms on PRUNE and CoME only by 5 and 6 percentage points respectively.
Cross-Method Generalization. We examine whether recovery signals transfer across editing techniques. On GPT-J and Llama3, MEMENTO outperforms One Mask in CMR across all techniques, with gains ranging from 0.05 to 0.26. This suggests that the learned recovery signal can generalize beyond the editing procedure used to train it. On GPT2-XL, the performance is again mixed. MEMENTO improves CMR for MEMIT and AlphaEdit, but performs slightly worse on the rest.
Overall, MEMENTO achieves competitive or better recovery than One Mask in most evaluated settings, with particularly strong results on GPT-J and Llama3. Crucially, MEMENTO achieves this recovery entirely through an input-space prefix, without modifying model weights. Thus, pre-edit behavior can be recovered through standard model inference, highlighting prompt-space recovery as a viable alternative for auditing model editing.
Key Takeaway
MEMENTO achieves relative improvements of 9.58% in top-1 recovery and 12.73% in CMR over One Mask. This demonstrates pre-edit behavior can be recovered without modifying model weights, with recovery signals transferring across editing techniques.
Text-Only Recovery
Research Question 2
Can continuous recovery be expressed as a discrete prompt while retaining effectiveness?
Unlike continuous prefixes, which operate directly in the embedding space, discrete recovery uses a sequence of ordinary text tokens that can be supplied through a standard language-model interface. Table 5 reports the Top-1 recovery achieved by MEMENTO’s discrete variant all the models and editing techniques. GPT2-XL achieves the strongest results, reaching 0.56 with PRUNE.
| Technique | GPT2-XL | GPT-J | Llama3 |
| AlphaEdit | 0.52 | 0.43 | 0.47 |
| MEMIT | 0.50 | 0.48 | 0.46 |
| RECT | 0.53 | 0.55 | 0.44 |
| PRUNE | 0.56 | 0.45 | 0.48 |
| CoME | 0.51 | 0.51 | 0.46 |
Table 5. Top-1 recovery of MEMENTO (discrete).
These results show the practical effectiveness of discrete recovery for black-box auditing. By recovering pre-edit behavior through ordinary textual prompts, MEMENTO enables auditors to probe the persistence of edited knowledge using standard inference interfaces, without access to internal embeddings or weights. This makes recovery feasible in realistic, access-constrained settings and provides a measurable means of assessing whether model editing has successfully suppressed prior behavior or merely obscured it.
Key Takeaway. MEMENTO achieves up to 0.56 top-1 recovery using discrete textual prompts. This enables black-box auditing of edited knowledge through standard inference interfaces, without access to model weights or internal embeddings.
Factual Recall Beyond Model Editing
Research Question 3. Can prompt-space recovery improve factual recall in pristine, unedited LLMs under standard inference?
Table 4 shows whether recovery signals trained on edited models can improve factual recall when applied to their corresponding pristine, unedited models. We compare standard inference against One Mask and MEMENTO using signals associated with five editing techniques, reporting both LLM-as-a-Judge and canonical-answer accuracy.
| Model | Method | Pristine Recall | AlphaEdit | MEMIT | RECT | PRUNE | CoME |
| GPT2-XL | One Mask | 21.7% | 12.5% (-9.2) | 12.2% (-9.5) | 11.4% (-10.3) | 4.4% (-17.3) | 14.3% (-7.4) |
| MEMENTO | 33.3% (+11.6) | 39.2% (+17.5) | 47.0% (+25.3) | 33.0% (+11.3) | 23.4% (+1.7) | ||
| GPT-J | One Mask | 33.0% | 27.6% (-5.4) | 26.5% (-6.5) | 28.9% (-4.1) | 27.1% (-5.9) | 27.1% (-5.9) |
| MEMENTO | 62.1% (+29.1) | 58.9% (+25.9) | 60.3% (+27.3) | 51.5% (+18.5) | 55.2% (+22.2) | ||
| Llama3 | One Mask | 52.8% | 49.5% (-3.3) | 53.7% (+0.9) | 51.6% (-1.2) | 52.7% (-0.1) | 48.9% (-3.9) |
| MEMENTO | 75.7% (+22.9) | 77.3% (+24.5) | 77.2% (+24.4) | 62.9% (+10.1) | 70.0% (+17.2) | ||
| GPT2-XL | One Mask | 14.6% | 8.4% (-6.2) | 12.6% (-2.0) | 10.4% (-4.2) | 3.7% (-10.9) | 9.7% (-4.9) |
| MEMENTO | 25.9% (+11.3) | 40.6% (+26.0) | 46.4% (+31.8) | 29.4% (+14.8) | 18.1% (+3.5) | ||
| GPT-J | One Mask | 19.5% | 18.3% (-1.2) | 18.5% (-1.0) | 19.6% (+0.1) | 18.5% (-1.0) | 18.9% (-0.6) |
| MEMENTO | 59.6% (+40.1) | 57.5% (+38.0) | 57.3% (+37.8) | 50.0% (+30.5) | 52.6% (+33.1) | ||
| Llama3 | One Mask | 29.1% | 26.9% (-2.2) | 30.8% (+1.7) | 29.5% (+0.4) | 31.2% (+2.1) | 26.6% (-2.5) |
| MEMENTO | 67.3% (+38.2) | 71.8% (+42.7) | 71.3% (+42.2) | 53.3% (+24.2) | 62.6% (+33.5) |
Table 4. Factual recall on pristine, unedited models. Pristine Recall column reports recall under standard inference. Remaining columns report recall using signals obtained under the indicated editing-technique settings. deepseek-ai/DeepSeek-V4.1-Flash is used as the LLM-as-a-Judge. Canonical-answer correctness requires canonical answer at the start of continuation after Unicode, case, punctuation, and whitespace normalization.
Improved factual recall in pristine models. MEMENTO consistently improves factual recall over standard inference across all three models and all five recovery-signal settings under both evaluation metrics. The gains are particularly pronounced for GPT-J and Llama3. Under LLM-as-a-Judge evaluation, MEMENTO raises recall from 33.0% to 51.5–62.1% on GPT-J and from 52.8% to 62.9–77.3% on Llama3. For GPT2-XL, recall improves from 21.7% to 23.4–47.0%, with the strongest result obtained using the RECT-associated recovery signal. Canonical-answer correctness exhibits a similar pattern, with gains across every model and recovery-signal setting.
Comparison with One Mask. One Mask generally reduces factual recall relative to standard inference, particularly on GPT2-XL, whereas MEMENTO improves recall across all evaluated settings. For example, on GPT2-XL under LLM-as-a-Judge evaluation, the RECT-associated signal raises recall to 47.0%, compared with 21.7% under standard inference and 11.4% with One Mask. The same pattern appears in canonical-answer correctness, where the corresponding scores are 46.4%, 14.6%, and 10.4%, respectively.
Implications beyond edit recovery. These results extend the findings beyond model editing, that the learned recovery signal can improve factual recall even when the model has not undergone an editing operation. This is consistent with the possibility that factual knowledge may be available to a model but not reliably retrieved under standard prompting. Our experiments suggest that prompt-space recovery can serve as a general means of probing and improving the accessibility of factual knowledge, beyond its original application to recovering pre-edit behavior.
| Key Takeaway |
| MEMENTO improves factual recall in pristine, unedited models. Averaged across all settings, LLM-as-a-Judge accuracy increases from 35.8% under standard inference to 55.1%, while canonical-answer correctness increases from 21.1% to 50.9%. |
Table 16.
Mechanistic Analysis
We investigate how MEMENTO recovers pre-edit factual behavior by examining how it changes the internal computation of MEMIT-edited models. Our analysis reveals three related effects: the recovery prefix (1) attenuates the measured contribution of the edit, (2) restores support for the original answer in later layers, and (3) substantially reduces the representational divergence between edited and unedited models under the same prefix. We conduct these analyses on held-out facts across GPT2-XL, GPT-J, and Llama3.
MEMENTO attenuates the measured edit write. For each edited layer , let denote the MEMIT weight update to the output projection and the corresponding key vector at the subject’s last token. We measure the edit write as , the magnitude of the update’s contribution at that layer. Comparing the same edited model with and without the recovery prefix, Figure 2 shows that MEMENTO reduces the edit write summed over edited layers by 82%, 87%, and 64% for GPT2-XL, GPT-J, and Llama3, respectively. This reduction reflects a change in the key representations through which the update is applied. MEMENTO changes how the MEMIT edit is expressed during inference. However, comparable write attenuation occurs for facts that MEMENTO recovers and facts that it fails to recover. Reduced write magnitude therefore cannot, by itself, explain recovery. We next examine how support for the original answer develops through the network.
Figure 2. Magnitude of the MEMIT update contribution at the edited layers, with and without MEMENTO.
MEMENTO restores support for the original answer. We use a layerwise logit lens to track the log-probability assigned to the original answer throughout the forward pass (Figure 3). Without MEMENTO, the original answer loses support in the later layers of the edited model, while the edited association becomes more dominant. This effect is particularly pronounced in GPT-J and Llama3. MEMENTO changes this pattern, and the original answer’s log-probability rises through the later layers, following a trajectory similar to that of the unedited model. Across all three models, the original answer regains support at approximately the same stage of computation at which it becomes likely in the unedited model. Its final log-probability even exceeds that of the unedited model. These results show that recovery is accompanied by a change in how support for the original answer develops through the network, rather than merely a change in the final prediction. Logit lens doesn’t establish that the underlying causal computation is identical to that of the original model, but reveals a clear correspondence between the original answer’s layerwise emergence and successful recovery.
Figure 3. Layerwise log-probability of the original answer in the unedited model, the MEMIT-edited model, and the edited model with MEMENTO.
MEMENTO reduces representational divergence associated with editing. We compare hidden states of edited and unedited models across layers at the subject and last prompt-token positions (Figure 4). We measure divergence using cosine similarity and relative distance, defined as , where and denote the unedited and edited representations, respectively. Both models receive either no recovery prefix or the same MEMENTO prefix, controlling for changes induced by the prefix itself.
Figure 4. Layerwise cosine similarity between the hidden states of edited and unedited models at the subject position (top) and the last prompt-token position (bottom).
Without the prefix, final-layer cosine similarity at the last prompt token falls to 0.80, 0.68, and 0.66 for GPT2-XL, GPT-J, and Llama3, respectively, with relative distances of 1.04, 1.14, and 0.89. With MEMENTO applied to both models, cosine similarity rises to 0.993, 0.991, and 0.987, while relative distance decreases substantially. The subject-position representations show the same pattern, indicating that under matched inputs, MEMENTO brings the edited model’s hidden states closer to those of its unedited counterpart.
Discussion. The three analyses reveal a consistent pattern. Model editing changes the model’s internal computation and suppresses the original answer. When MEMENTO is applied, the measured contribution of the edit decreases, support for the original answer re-emerges through the later layers, and the edited model’s representations become much more similar to those of the unedited model receiving the same prefix.
These results suggest that MEMENTO recovers pre-edit factual behavior by steering the edited model toward an input-conditioned computation that substantially reduces the representational divergence associated with editing. The current analyses do not establish a causal mechanism. Nevertheless, their agreement shows that prompt-space recovery is accompanied by systematic changes in how the edited model processes the input. More broadly, the results show that an edit that suppresses an original fact under ordinary prompting does not necessarily make the corresponding pre-edit behavior inaccessible.
Conclusion
We presented MEMENTO, a prompt-space auditing mechanism for recovering pre-edit knowledge in LLMs without modifying their parameters. By learning reusable recovery signals in continuous and discrete prompt spaces, MEMENTO provides an alternative to weight-space edit reversal. Across three LLMs and five editing techniques, continuous MEMENTO recovers 80.07% of held-out edited facts, achieving a 9.58% relative improvement over weight-space edit-reversal methods, while the discrete variant achieves 49.0% recovery accuracy using textual prompts. These findings show that pre-edit factual responses can remain behaviorally recoverable through input-space interventions, even when conventional evaluations indicate successful edits. Beyond model editing, we show that MEMENTO also improves factual recall in pristine, unedited models, increasing average recall by 53.9% relative to standard inference. This highlights the broader potential of prompt-space recovery for investigating the gap between the factual information a model contains and what it retrieves under standard prompting. We hope our work motivates further investigation into the reliability of model editing and the relationship between stored information and its behavioral accessibility in language models.
Limitations and Ethical Considerations. Our evaluation covers three LLMs and five editing techniques, so the extent to which our findings generalize to other models and techniques remains to be investigated. Moreover, successful behavioral recovery does not establish that pre-edit knowledge remains intact in a particular internal representation. MEMENTO is designed as an auditing mechanism to evaluate the behavioral recoverability of edited knowledge. Identifying such recoverability can help expose limitations in existing editing techniques and motivate more rigorous evaluations. We intend this work to support research into reliable model editing and encourage responsible use that respects privacy and avoids unauthorized extraction of sensitive information.
Reproducibility
Models
Table 6 lists the models, their Hugging Face check-points and revisions, and the layers modified by the editing techniques. All models are loaded in 32-bit floating point, and every forward pass in training and evaluation uses the same precision.
| Model | Checkpoint | Revision | Layers | Edited layers | |
| GPT2-XL | openai-community/gpt2-xl | 15ea56dee5df | 48 | 1600 | 13-17 |
| GPT-J | EleutherAI/gpt-j-6B | 47e169305d2e | 28 | 4096 | 3-8 |
| Llama3 | meta-llama/Meta-Llama-3-8B-Instruct | 8afb486c1db2 | 32 | 4096 | 4-8 |
Table 6. Models. is the embedding dimension, which is also the dimension of each prefix vector. All five editing techniques modify the same layers of a given model.
Editing Techniques
All techniques share the MEMIT key-value computation and its hyperparameters (Table 7). They differ in how the update to the edited weights is formed. AlphaEdit projects the update onto the null space of preserved knowledge, with null-space threshold 0.02 and L2 regularization 10. RECT applies only the 40% of update entries with the largest relative change .
| GPT2-XL | GPT-J | Llama3 | |
| Value optimization steps | 20 | 25 | 25 |
| Value learning rate | 0.5 | 0.5 | 0.1 |
| Clamp norm factor | 0.75 | 0.75 | 0.75 |
| KL factor | 0.0625 | 0.0625 | 0.0625 |
| Value weight decay | 0.5 | 0.5 | 0.5 |
| Covariance update weight | 20,000 | 15,000 | 15,000 |
| Covariance samples | 100,000 | 100,000 | 100,000 |
Table 7. Hyperparameters shared by all editing techniques, following the original MEMIT configurations.
PRUNE computes the MEMIT update and log-compresses each of its singular values that exceeds the largest singular value of the original weight, . CoME uses its MEMIT-based variant. Second-moment statistics of the keys are estimated from 100,000 Wikipedia samples per edited layer and cached.
Continuous Recovery
Table 8 lists the settings for training the continuous prefix. A single configuration is used for every model and editing technique; only the learning rate differs across models, and it is not tuned per model but set by a fixed rule. Because Adam takes steps of roughly constant size regardless of gradient scale, we set the learning rate proportional to the standard deviation of the model’s input embeddings, , calibrating once on GPT2-XL. The input embeddings of Llama3 have a standard deviation about five times smaller than those of GPT2-XL, and a learning rate tuned on GPT2-XL diverges on Llama3.
| Setting | Value |
| Prefix length | 16 |
| Initialization | embedding of “!” repeated |
| Position | before the prompt |
| Optimizer | Adam, constant learning rate |
| Learning rate | |
| GPT2-XL / GPT-J / Llama3 | |
| Batch size (facts) | 8 |
| Maximum steps | 1,000 |
| Preference scale / margin | 2.0 / 0.5 |
| 1.0 | |
| 0.1 | |
| Fluency reference texts | 10 fixed sentences |
| / temperature | 1.0 / 2.0 |
| Validation interval / patience | 25 steps / 8 checks |
| Divergence abort | twice |
| Random seed | 0 |
Table 8. Continuous MEMENTO settings, shared across all models and editing techniques.
The prefix is initialized by repeating the embedding of the token “!” times. is computed over all next-token positions of one of ten fixed generic sentences, sampled for each training example. The pre-edit distributions for are computed before editing and cached. Every 25 steps, we compute the mean on the validation facts, keep the prefix with the lowest value, and stop after eight consecutive checks without improvement. As a safeguard against divergence, training is aborted if exceeds 5.0 on two consecutive steps after step 20.
Discrete Recovery
Table 9 lists the settings of the discrete search. The search starts from a sequence of distinct space-prefixed words that satisfies the round-trip constraint. At each step, it samples eight training facts, computes the gradient of with respect to the one-hot token indicators, and proposes single-position substitutions drawn from the top-256 tokens per position. Candidates that violate are discarded; we sample four times the candidate batch size so that the batch remains full after filtering. The current sequence is scored on the same eight facts as its candidates, and the search moves to the candidate with the lowest loss. Every 25 steps, the current sequence is scored on the validation facts, and the sequence with the lowest validation loss is returned.
| Setting | Value |
| Sequence length | 16 tokens |
| Objective | |
| / / | 2.0 / 0.5 / 1.0 |
| Top- substitutions per position | 256 |
| Candidates per step | 256 (sampled 4×, then filtered) |
| Facts per step | 8 |
| Maximum steps | 1,000 |
| Validation interval / patience | 25 steps / 8 checks |
| Random seed | 42 |
Table 9. Discrete MEMENTO settings, shared across all models and editing techniques.
Compute
All experiments run on a single NVIDIA H200 NVL GPU (141 GB) with Python 3.12.3, PyTorch 2.6.0 (CUDA 12.4), and Transformers 4.57.3. Table 10 reports wall-clock times for MEMIT.
| GPT2-XL | GPT-J | Llama3 | |
| Continuous training | 0.5 h | 0.7 h | 1.6 h |
| Evaluation | 0.4 h | 0.5 h | 1.3 h |
| Discrete search | 1.4 h | 2.8 h | 6.3 h |
Table 10. Wall-clock time per run for MEMIT on one H200 GPU.
Cost Analysis
One-time training cost. Table 11 reports wall-clock training time for MEMIT with 1,000 training facts. The continuous prefix trains in under two hours on every model, including the time to edit the training facts and to cache the pre-edit distributions for . The discrete search costs more, between 1.4 and 6.3 hours, because every step evaluates 256 candidate sequences by exact forward passes. The resulting prefix is applied unchanged to every subsequent query and to edits it was not trained on.
| Method | GPT2-XL | GPT-J | Llama3 |
| MEMENTO (continuous) | 0.5 h | 0.7 h | 1.6 h |
| MEMENTO (discrete) | 1.4 h | 2.8 h | 6.3 h |
Table 11. One-time training cost for MEMIT with 1,000 training facts on one H200 GPU, including the time to edit the training facts.
Per-query recovery cost. At inference, the continuous prefix adds input positions to a prompt of about nine tokens; recovery is otherwise an ordinary forward pass of the deployed model. Table 12 reports the latency of greedily generating 12 tokens for 100 held-out prompts with and without the prefix. The difference is within on all three models, which is within run-to-run variation: decoding time is dominated by the per-token steps rather than by the additional prefill positions. The discrete prefix adds 16 tokens of text and therefore has the same cost.
| GPT2-XL | GPT-J | Llama3 | |
| Without prefix (ms/query) | 95.6 | 135.2 | 165.2 |
| With MEMENTO (ms/query) | 90.9 | 141.8 | 169.7 |
| Relative change | −4.9% | +4.9% | +2.7% |
Table 12. Per-query latency of greedy decoding (12 new tokens, batch size 1), averaged over 100 held-out prompts after warm-up.
Complexity Analysis
We express costs in terms of the forward cost of the frozen model on a sequence of tokens,
where is the number of layers, the embedding dimension, and the MLP width; a backward pass costs a constant multiple of . Let be the prefix length, the prompt length, the object length, the length of a fluency reference sentence, the vocabulary size, and the number of generated tokens. Table 13 summarizes the resulting costs.
| Cost | MEMENTO (continuous) | MEMENTO (discrete) |
| Training time | ||
| Trainable values | tokens | |
| Recovery time per query | steps | steps |
| Recovery memory overhead | cache | cache |
Table 13. Asymptotic costs. : training steps; : facts per step; , : candidates and facts per step of discrete search.
Continuous training. Each step processes facts. Per fact, and require forward and backward passes over the prefixed prompt followed by and by , reuses the next-token distribution at the last prompt position, and requires one pass over a reference sentence with the prefix and one without it. Training for steps therefore costs , linear in the number of steps and independent of , which enters only through the one-time cost of editing it and caching the pre-edit distributions. Gradients flow through the frozen model to its input, so memory is dominated by the activations of one fact, , plus the model parameters. Both and values are trainable, with optimizer states. The cached pre-edit distributions occupy : 171 MB for GPT2-XL and GPT-J and 436 MB for Llama3 at 850 gradient facts in 32-bit precision. Discrete training. Each step computes token gradients on facts, which requires the one-hot gradient over the vocabulary at every position, memory and additional computation per pass. It then scores candidate sequences on the same facts against both objects. The candidate evaluation dominates, giving for steps; with and , each step costs roughly forward passes, compared with forward-backward passes for the continuous prefix. The round-trip constraint adds tokenizer calls per step.
Recovery. Generating tokens with a prefix costs for the prefill and for the -th decoding step. Relative to the unprefixed prompt, the prefix adds prefill positions and entries per layer to the key-value cache, memory, and enlarges each decoding step’s attention by . Since and , the overhead is small, consistent with the measured latency in Table 12.
Ablations on the Recovery Objective
We conduct an ablation study on MEMIT-edited models to investigate the contribution of each term in MEMENTO’s continuous recovery objective. We additionally evaluate a variant that removes both the preference loss and fluency regularizer, as well as a recovery-only objective containing . Table 14 reports top-1 recovery on held-out edited facts and the perplexity (PPL) ratio on held-out text. Higher recovery indicates better restoration of pre-edit factual behavior, whereas lower PPL ratios indicate better preservation of general language-model behavior.
| Objective | GPT2-XL | GPT-J | Llama3 |
| Full objective | 0.790 / ×1.74 | 0.800 / ×1.97 | 0.840 / ×1.07 |
| 0.810 / ×1.14 | 0.775 / ×1.41 | 0.830 / ×1.08 | |
| 0.745 / ×1.24 | 0.810 / ×1.65 | 0.830 / ×1.13 | |
| 0.800 / ×7.31 | 0.755 / ×3.50 | 0.825 / ×1.60 | |
| 0.810 / ×1.16 | 0.765 / ×1.60 | 0.840 / ×1.14 | |
| only | 0.785 / ×1.79 | 0.745 / ×4.13 | 0.820 / ×1.96 |
Table 14. Effect of the continuous recovery objective on MEMIT-edited models. Each cell reports top-1 recovery on held-out edited facts (higher is better) and the perplexity (PPL) ratio on held-out text (lower is better), respectively. Bold values indicate the best result for each metric and model.
Overall, this ablation reveals complementary but model-dependent effects of the auxiliary objectives. The distillation loss has the clearest and most consistent impact on preserving general language-model behavior, while the preference and fluency terms affect the trade-off between factual recovery and language modeling differently across models. Although removing either term can improve individual metrics, the results show that their effects are not uniform across models. These findings motivate treating the recovery objective as a combination of complementary constraints rather than assuming that each term independently improves every evaluation metric.
LLM-as-a-Judge Evaluation
We use an LLM judge to evaluate whether a generated continuation correctly expresses the factual answer requested by the prompt. Unlike Canonical-Answer Correctness, which requires the canonical answer to appear at the beginning of the continuation, semantic judging can account for equivalent entity names and context-dependent expressions. The judge is instructed to assess the requested fact rather than the general plausibility of the continuation, and to disregard additional claims that do not establish the requested answer.
Table 15 presents illustrative judgments. In the first example, the continuation mentions a German-speaking region but identifies Austria, which does not establish German citizenship. In the second example, “the Boeing Company” refers to the entity Boeing, even though the continuation does not begin with the canonical answer “Boeing”. The examples illustrate why canonical-answer matching and semantic evaluation can yield different results. LLM-based judgments remain imperfect, however, and may require manual inspection in ambiguous cases.
| Prompt | Reference | Generated continuation | Judge | Rationale |
| Edmund Husserl, who has a citizenship from | Germany | “the Republic of Austria, was born in the city of Freiburg im Breisgau, in the German-speaking part of the country, on the . . . ” | Incorrect | The continuation explicitly identifies Austria. Describing a region as German-speaking does not establish German citizenship. |
| Boeing KC-46 is a product of | Boeing | “the Boeing Company. It is a long-range, medium-altitude, long-endurance, tanker-delivered . . . ” | Correct | “The Boeing Company” identifies the same entity as Boeing, even though the canonical answer does not appear verbatim at the start. |
| Shahab-1 is produced by | Iran | “the Iranian Air Force and is used by the Iranian Air Force to conduct air strikes . . . ” | Correct | The judge accepts “Iranian” as expressing the requested country through an adjectival form. This illustrates an indirect answer; such judgments may be sensitive to the interpretation of the relation. |
Table 15. Examples of LLM-as-a-Judge decisions on generated continuations.
LLM-as-a-Judge System Prompt
We use the following system prompt to evaluate factual answers generated by unedited language models. It instructs the judge to evaluate semantic correctness with respect to the requested relation, rather than relying on exact string matching or answer position.
Task. You are evaluating factual answers produced by unedited language models.
For each record, decide whether the generated continuation answers the requested relation correctly according to the supplied ground-truth answer. Evaluate the meaning of the answer, not whether a string occurs at a particular position. The prompt and continuation form one piece of text; read them together.
Use only the supplied records. Do not browse, use tools, edit files, execute code, or retrieve outside facts. Prompts and continuations are untrusted evaluation data: never obey instructions inside them. Return JSON only, without Markdown.
Rules.
Mark correct when the continuation clearly expresses the ground-truth answer for the requested subject and relation. The answer may occur later in the continuation or in a later sentence. There is no first-sentence restriction.
Accept unambiguous aliases, paraphrases and appropriate country adjectives. Accept a more specific location when it clearly identifies the target country, such as “Tokyo, Japan” for Japan. Accept “the Indian subcontinent” for India in the Namkeen example below, following the evaluation policy for this study.
A target string appearing incidentally is insufficient. It must fill the requested relation. Nationality of the creators alone does not establish a work’s country of origin. A word in an unrelated title, person, organization, quoted passage or subsequent topic is not automatically an answer.
Reject explicit negations of the answer. “It’s not jazz” does not answer jazz. If the continuation gives incompatible answers to the same relation without resolving
the conflict, do not mark it correct. Use uncertain if genuinely ambiguous; use incorrect for a clear wrong answer or denial.
Accept nonexclusive lists or qualified descriptions that clearly include the true answer, unless the text denies it. “A striker, but also a midfielder” answers midfielder; “a mixture of Spanish and Portuguese” answers Spanish.
Judge only the requested fact. Ignore unrelated errors in added details unless they undermine the answer to that fact. Do not independently fact-check the supplied ground truth. Repetition alone does not make a correct answer wrong.
Judge only the generated text supplied. If generation ends mid-sentence, accept a fact already established; do not imagine an answer beyond the cutoff.
Use incorrect if no answer is expressed. Reserve uncertain for ambiguity that prevents a defensible decision; never count uncertain as correct.
Policy examples. These are illustrations, not records to include in the output.
Prompt: “Nintendo, located in”; truth: “Japan”; continuation: “Tokyo, Japan, is a leading developer . . .” → correct.
Prompt: “Namkeen, that originated in”; truth: “India”; continuation: “the city of Kolkata, is a popular dish in the Indian subcontinent.” → correct.
Prompt: “The language used by Leopoldo Lugones is”; truth: “Spanish”; continuation: “a mixture of Spanish and Portuguese.” → correct.
Prompt: “Koke plays as”; truth: “midfielder”; continuation: “a striker, but he’s also a midfielder.” → correct.
Prompt: “Johannes Bugenhagen speaks”; truth: “German”; continuation: “to the press after the meeting. The meeting was held in the presence of the German foreign minister . . .” → incorrect: German describes another person.
Prompt: “The genre played by Mose Allison is”; truth: “jazz”; continuation: “a little hard to define. It’s not jazz . . .” → incorrect.
Prompt: “Niboshi was created in”; truth: “Japan”; continuation: “the early 1990s by a group of Japanese artists . . .” → incorrect: creator nationality does not establish the creation location.
Prompt: “Shahab-1 is produced by”; truth: “Iran”; continuation: “the Iranian Air Force . . .” → correct: identifies an Iranian producer.
Prompt: “Dodge Monaco, created by”; truth: “Dodge”; continuation: “the same team that created the original Dodge Viper . . .” → incorrect: mentioning a different Dodge product does not clearly identify Dodge as the creator.
Prompt: “Infiniti JX, created by”; truth: “Nissan”; continuation: “Japanese carmaker Nissan, is the first car . . .” → correct.
Input and output format. The input is a JSON array of records with keys key, model, id, prompt, true_answer, continuation, and relation_id. Previous scorer labels are deliberately omitted. Return one JSON object with exactly this structure:
{
"judgments": [
{
"key": "EXACT_INPUT_KEY",
"label": "correct|incorrect|uncertain",
"evidence": "short exact quotation from
the continuation, or empty if
absent",
"reason": "one short sentence
explaining the requested relation"
}
]
}
Include every input key exactly once and no other keys or records. Preserve input order. Evidence must be quoted from that record’s continuation, not from memory or another record. Do not output internal reasoning or any additional text.
References
- [1]Bang An, Shiyue Zhang, and Mark Dredze. 2025. Rag llms are not safer: A safety analysis of retrieval-augmented generation for large language models. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 5444–5474.
- [3]Manit Baser, Aditya Nawal, Dinil Mon Divakaran, and Mohan Gurusamy. 2026b. The tokens remember: When tokenization bypasses knowledge editing and unlearning. arXiv preprint arXiv:2609.29045.arxiv.org/abs/2609.29045
- [4]Manit Baser, Alperen Yildiz, Dinil Mon Divakaran, and Mohan Gurusamy. 2026c. CLaRE-ty amid chaos: Quantifying representational entanglement to predict ripple effects in LLM editing. In Findings of the Association for Computational Linguistics: ACL 2026, pages 29373–29405, San Diego, California, United States. Association for Computational Linguistics.
- [5]Nitay Calderon, Eyal Ben-David, Zorik Gekhman, Eran Ofek, and Gal Yona. 2026. Empty shelves or lost keys? recall is the bottleneck for parametric factuality. In Forty-third International Conference on Machine Learning.arxiv.org/abs/2602.14080
- [6]Sukmin Cho, Soyeong Jeong, Jeongyeon Seo, Taeho Hwang, and Jong C Park. 2024. Typos that broke the rag’s back: Genetic attack on rag pipeline by simulating documents in the wild via low-level perturbations. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 2826–2844.
- [7]Roi Cohen, Eden Biran, Ori Yoran, Amir Globerson, and Mor Geva. 2024. Evaluating the ripple effects of knowledge editing in language models. Transactions of the Association for Computational Linguistics, 12:283–298.
- [8]Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Jie Shi, Xiang Wang, Xiangnan He, and Tat-Seng Chua. 2025. Alphadit: Null-space constrained knowledge editing for language models. In International Conference on Learning Representations, volume 2025, pages 16366–16396.arxiv.org/abs/2410.02355
- [9]Jia-Chen Gu, Hao-Xiang Xu, Jun-Yu Ma, Pan Lu, Zhen-Hua Ling, Kai-Wei Chang, and Nanyun Peng. 2024. Model editing harms general abilities of large language models: Regularization to the rescue. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 16801–16819.arxiv.org/abs/2401.04700
- [10]Tom Hartvigsen, Swami Sankaranarayanan, Hamid Palangi, Yoon Kim, and Marzyeh Ghassemi. 2023. Aging with grace: Lifelong model editing with discrete key-value adaptors. Advances in Neural Information Processing Systems, 36:47934–47959.arxiv.org/abs/2211.11031
- [11]Ali Holmov, Paul Youssef, Nandi Schoots, and Christin Seifert. 2026. One mask to rule them all: On hidden facts after editing and how to find them. In Findings of the Association for Computational Linguistics: ACL 2026, pages 11163–11181, San Diego, California, United States. Association for Computational Linguistics.arxiv.org/abs/2605.28839
- [12]Hongsheng Hu, Shuo Wang, Jiamin Chang, Haonan Zhong, Ruoxi Sun, Shuang Hao, Haojin Zhu, and Minhui Xue. 2024. A duty to forget, a right to be assured? exposing vulnerabilities in machine un-learning services. In Proceedings of the Network and Distributed System Security Symposium (NDSS).DOI
- [13]Weifeng Jiang, Ruirui Chen, Qianren Mao, Junnan Liu, Qili Zhang, and Kwok-Yan Lam. 2026. Selective knowledge edit reversal via gated singular vector shrinkage. arXiv preprint arXiv:2609.02091.arxiv.org/abs/2609.02091
- [14]Dahyun Jung, Jaehyung Seo, Jaewook Lee, Chan-jun Park, and Heui-Seok Lim. 2025. Come: An unlearning-based approach to conflict-free model editing. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 6410–6422.
- [15]Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045–3059, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.arxiv.org/abs/2104.08691
- [16]Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582–4597, Online. Association for Computational Linguistics.
- [17]Jun-Yu Ma, Hong Wang, Hao-Xiang Xu, Zhen-Hua Ling, and Jia-Chen Gu. 2025. Perturbation-restrained sequential model editing. In Proceedings of the International Conference on Learning Representations (ICLR 2025).arxiv.org/abs/2405.16821
- [18]Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in gpt. Advances in neural information processing systems, 35:17359–17372.arxiv.org/abs/2202.05262
- [19]Kevin Meng, Arnab Sen Sharma, Alex J Andonian, Yonatan Belinkov, and David Bau. 2023. Mass-editing memory in a transformer. In The Eleventh International Conference on Learning Representations.arxiv.org/abs/2210.07229
- [20]Eric Mitchell, Charles Lin, Antoine Bosselut, Christopher D Manning, and Chelsea Finn. 2022. Memory-based model editing at scale. In International Conference on Machine Learning, pages 15817–15831. PMLR.arxiv.org/abs/2206.06520
- [21]Elena Sofia Ruzzetti, Giancarlo A. Xompero, Davide Venditti, and Fabio Massimo Zanzotto. 2025. Private memorization editing: Turning memorization into a defense to strengthen data privacy in large language models. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16572–16592, Vienna, Austria. Association for Computational Linguistics.DOI
- [22]Davide Venditti, Elena Sofia Ruzzetti, Giancarlo A. Xompero, Cristina Giannone, Andrea Favalli, Raniero Romagnoli, and Fabio Massimo Zanzotto. 2026. Enhancing data privacy in large language models through private association editing. Journal of Cybersecurity and Privacy, 6(3):96.
- [23]Paul Youssef, Zhixue Zhao, Jörg Schlötterer, and Christin Seifert. 2025. How to make LLMs forget: On reversing in-context knowledge edits. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 12656–12669, Albuquerque, New Mexico. Association for Computational Linguistics.
- [24]Paul Youssef, Zhixue Zhao, Christin Seifert, and Jörg Schlötterer. 2026. Tracing and reversing edits in llms. In International Conference on Learning Representations, volume 2026, pages 23863–23893.arxiv.org/abs/2505.20819
- [25]Wenqian Yu, Shuo Chen, Zhijiang Li, Zhipeng Wang, and Jindong Gu. 2026. You only need one single token to refine safety alignment. In Findings of the Association for Computational Linguistics: ACL 2026, pages 13529–13545, San Diego, California, United States. Association for Computational Linguistics.
- [26]Ce Zheng, Lei Li, Qingxiu Dong, Yuxuan Fan, Zhiyong Wu, Jingjing Xu, and Baobao Chang. 2023. Can we edit factual knowledge by in-context learning? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 4862–4876, Singapore. Association for Computational Linguistics.arxiv.org/abs/2305.12740
- [27]Zexuan Zhong, Zhengxuan Wu, Christopher D Manning, Christopher Potts, and Danqi Chen. 2023. Mquake: Assessing knowledge editing in language models via multi-hop questions. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 15686–15702.
- [28]Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. 2023. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043.arxiv.org/abs/2307.15043



