acceptodds
Under review as a conference paper at ICLR 2027

Gradient as Behavior Fingerprint: Stable Sequential Model Editing via Gradient Drift Regularization

Abstract

Large language models (LLMs) have achieved remarkable success across a wide range of tasks, but outdated or incorrect knowledge may persist in their parameters, leading to hallucinations and factual errors. Model editing aims to address this issue by enabling targeted updates to specific knowledge. However, in sequential editing settings, repeated updates can interfere with one another, resulting in accumulated instability and degradation of previously edited knowledge. In this paper, we revisit sequential editing from a gradient-based perspective. We use the gradient of the target log-likelihood with respect to the editing residual as a target-conditioned descriptor of first-order local perturbation response. We establish a Taylor bound showing that the gradient discrepancy between two state-specific reference centers controls the difference between their centered local responses up to a second-order remainder. Building on this diagnostic, we propose Gradient Drift Regularization (GDR). For each historical edit, GDR stores a gradient anchor at its insertion-time reference center and later computes a query gradient from the current model state when that edit is selected. GDR uses the discrepancy between these gradients to construct a fixed, dimension-wise quadratic penalty on the current editing residual, discouraging updates along coordinates with larger observed historical discrepancies. To improve efficiency, we further introduce a preconditioned key-space selection mechanism that identifies historical edits potentially affected by the current update, limiting query-gradient evaluation to a top- subset. Experiments on LLaMA-3.1, GPT-2-XL, and Qwen-2.5 show that GDR preserves overall model performance and improves the retention of historical edits over long editing sequences. The results also reveal a strong empirical association between lower gradient drift and better historical retention.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Sequential Model Editing, Knowledge Editing, Gradient Drift Regularization, Knowledge Retention

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_qBDEOpenReview ↗
Rating
3(+1)
Confidence
1(−3)
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes Gradient Drift Regularization (GDR) to mitigate the "catastrophic forgetting" of previously acquired knowledge during sequential knowledge editing in language models. The method preserves the residual-space gradients from historical edits (captured at the time of insertion) and recomputes gradients for a subset of historical samples prior to subsequent edits. The squared coordinate-wise difference between these two sets of gradients is then used as a quadratic regularization weight for the current residual optimization. A top-k mechanism, based on covariance-preconditioned key interactions, is employed to constrain the computational cost of calculating historical gradients. The authors derive a Taylor bound for the centered local response difference and demonstrate strong performance in both knowledge editing and the retention of previously learned information across three different models and benchmarks such as CounterFact and ZsRE.

Critical strengths
  1. Clear design addressing a practical problem. The issue of cumulative interference during sequential editing is a significant area of ​​study. By leveraging the discrepancy between historical reference states and the current state, GDR provides a regularization signal that differs from standard single-state importance estimation (Section 5).
  2. Ease of integration and well-defined computational bounds. The weights defined in Equation 15 remain fixed during residual optimization, and the top-k mechanism limits the overhead of computing historical gradients. Appendix D.6 transparently acknowledges that both the scanning of historical samples and the storage requirements scale with the number of edits performed.
  3. Targeted experimental comparisons. Tables 1 and 2 cover three models and two editing datasets, including comparisons against baselines such as AlphaEdit and HSE. Tables 4 and 6 provide further analysis of selection strategies, response signals, and weight transformations. Notably, for the LLaMA model on the CounterFact generalization task, the performance improves from 78.15 (AlphaEdit) to 93.95—a significant gain, provided it can be verified.
Critical weaknesses
  1. Inconsistencies in statistical reporting. In Appendix C.3, error is defined as the instance-level Standard Error of the Mean (SEM), and binomial standard errors are provided for binary metrics; however, there are discrepancies between these definitions and some of the values ​​presented in the tables. If the "CounterFact efficacy" values ​​in Table 1 are derived from binary judgments for each of 1,000 individual edits, then the 97.50% reported for LLaMA GDR should correspond to a margin of error of approximately 0.494 percentage points (rather than the reported 0.28), and the 50.90% for MEMIT should correspond to approximately 1.581 (rather than 0.76). Appendix C.1 specifies that the RTE evaluation uses 277 samples; however, the figures of 78.30% and 68.12% in Table 7 cannot be obtained by rounding the result of an integer number of correct predictions divided by 277. While differences in sample units or averaging across multiple runs might account for some of these discrepancies, the text provides no clarification. This directly impacts the reproducibility of the results; furthermore, instance-level standard error of the mean (SEM) cannot serve as a substitute for accounting for the order of edits or the variance across experimental runs.
  2. The independent role of the gradient difference across states has not been sufficiently established. The perturbation term ϵ\epsilon (representing shared historical samples) in the theorem is not the same entity as the residual δ\delta (representing the current edit) in Equation 15; the mechanism by which parameter updates bridge these two remains undemonstrated. The "anchor" is derived from the model prior to the update plus an optimized residual, while the "query" is derived from the model after the update—the difference between them may also incorporate the initial update error. When the anchor gradient is small, the squared drift approximates the squared current gradient; the "anchor Fisher" data in Table 6 does not rule out this alternative explanation. Employing controls such as "query-gradient-only," using a post-update anchor, or implementing residual regularization (e.g., by matching update norms or shuffling coordinates) would provide a more direct test of the proposed mechanism. While the current results support an empirical correlation, they are insufficient to definitively distinguish between these potential explanations. Admittedly, this is a higher-level critique and not grounds for rejection; however, the paper would be strengthened if the authors could design experiments to further substantiate the robustness or scope of this empirical correlation.
  3. There is a correctable mathematical inconsistency in the formulation of the closed-form update. The exact interpolation condition in Eq. 4 generally does not follow from Eq. 5. With C=K₀K₀ᵀ≻0, r=v*−Wk, and t=kᵀC⁻¹k, Eq. 5 gives Δk=r·t/(1+t), rather than r. Deriving Eq. 29 from Eq. 27 also requires the background residual term to vanish, for example when V₀=WK₀; approximate satisfaction alone does not yield the stated equality. The paper should specify whether background values are refreshed and how the covariance scaling in C.3 enters these equations. Equation 61 also reuses the edited target for unrelated locality questions without defining their own correct answers.
  4. The scope of certain diagnostic interpretations lacks clarity. In Figures 2 and 6, the "earliest 20%" cohort varies with TT; keeping this early cohort fixed would be more appropriate for directly illustrating the trajectory of "forgetting." For Table 3, Figure 5, and the timing analysis in C.3, the specific datasets and the criteria used for defining the temporal boundaries should be specified. Furthermore, the "Edit Score" in Table 3 lacks an explicit formula.
Questions for authors

see weakness 1,3,4.

If you can prove the experience relationship in weakness 2 by experiments, I will raise my score.

AFTER REBUTTAL

Essentially, I would have liked to give the author a score of 2.5, as they partially substantiated some of the weaknesses I identified; however, given the limited strength of this substantiation, I have opted to assign a final score of 3 with a confidence level of 1.

6 comments
AuthorsOpenReview ↗

Rebuttal by Authors (Part 1 of 2)

Thank you very much for your careful review and constructive comments.

W1 (Inconsistencies in statistical reporting).

(1) Efficacy and SEM. CounterFact Eff. assigns one binary outcome per edit, with percentage-point SEM 100p^(1−p^)/N100\sqrt{\hat p(1-\hat p)/N}. The two LLaMA entries in Table 1 highlighted by the reviewer correspond to 975 and 509 successes out of 1,000 edits, respectively.

MethodOriginalCorrected
GDR97.50 ± 0.2897.50 ± 0.49
MEMIT [1]50.90 ± 0.7650.90 ± 1.58

We recomputed the CounterFact Eff. SEMs in Tables 1 and 3 (GPT-2-XL in Appendix D.1). Fractional or continuous scores, including ZsRE and the other CounterFact metrics, use s/Ns/\sqrt N, where ss is the sample SD across edit-level scores. ZsRE first averages next-token correctness given gold prefixes within each edit, then across edits. Section 6.1 and Appendices C.1 and C.3 now clarify these aggregation rules and SEM definitions.

(2) Downstream evaluation denominators and results. Generation or answer-parsing failures caused some examples to be skipped during aggregation, reducing the denominator. For GDR on RTE, the reported 68.12% came from 188 correct predictions among 276 scored examples after one example was omitted. Similar omissions affected GSM8K and SAMSum.

We added generation retries and parsing fallbacks, reran evaluation, and verified that every example was scored. On all 277 RTE examples, the corrected accuracies are 78.34% (217/277) before editing and 68.23% (189/277) after GDR. We updated Table 11 in Appendix D.4 (original Table 7) and clarified the evaluation splits in Appendix C.1. We apologize for these reporting errors and appreciate the reviewer’s careful checks.

(3) Variability across edit orders. To assess edit-order variability, we evaluated three shared random permutations of the same 1,000 CounterFact edits on LLaMA-3.1-8B-Instruct, resetting to the same initial checkpoint for each run. Table 10 summarizes the comparison.

MethodEff. ↑Gen. ↑Spe. ↑Hist. ↑
GDR97.37 ± 0.2193.90 ± 0.3066.37 ± 0.3191.17 ± 0.76
AlphaEdit [2]93.40 ± 0.6278.20 ± 0.9262.10 ± 0.6277.67 ± 0.76

These are means ± sample SD across the three orders, distinct from instance-level SEMs within one run. Hist. evaluates the first 200 edits of each permutation under the final model.

GDR scores higher on all four metrics in every tested order. Its ranges are 0.40, 0.60, 0.60, and 1.50 percentage points for Eff., Gen., Spe., and Hist., respectively. See Appendix D.3.4, Table 10, and Fig. 5 for per-order results.

W2 (Independent role of cross-state gradient differences).

(1) Connecting editing residuals to historical perturbations. In Theorem 1, ϵ\epsilon is a common local probe of historical representations. For a fixed historical key, Appendix B.2, Eqs. (40)–(42), gives Δvj=Δikj=αijri\Delta v_j=\Delta_i k_j=\alpha_{ij}r_i, where rir_i is the layerwise residual allocated through Eq. (21). For a single additive target-layer write with ri=δir_i=\delta_i, this yields ϵj=αijδi\epsilon_j=\alpha_{ij}\delta_i. The current residual induces the historical perturbation through δi→ΔWi→Δhj\delta_i\rightarrow\Delta W_i\rightarrow\Delta h_j.

Under its smoothness assumptions, Theorem 1 bounds differences between the centered responses to this common probe using gradient drift and curvature. Appendix B.3 further establishes ∑ej∈Bi⟨Δgj,δ⟩2≤d LGDR(i)(δ)\sum_{e_j\in\mathcal B_i}\langle\Delta g_j,\delta\rangle^2\leq d\,\mathcal L_{\mathrm{GDR}}^{(i)}(\delta) (Eq. (57)), where Δgj=gjquery−gjanc\Delta g_j=g_j^{\mathrm{query}}-g_j^{\mathrm{anc}} and d=dim⁡(δ)d=\dim(\delta). This provides a local diagonal surrogate for controlling first-order response differences. The controls in Part 2 assess actual multilayer writes.

(2) Anchor reference and the independent role of gradient drift. The original anchor is evaluated at the optimized-target reference, whereas the post-update anchor is evaluated on GjG_j immediately after the actual write (Eq. (74)). Eq. (78) separates insertion-time discrepancy from subsequent changes as gtquery−ganc=(gtquery−gpost)+(gpost−ganc)g_t^{\mathrm{query}}-g^{\mathrm{anc}}=(g_t^{\mathrm{query}}-g^{\mathrm{post}})+(g^{\mathrm{post}}-g^{\mathrm{anc}}).

On the fixed first 200 edits at T=1000T=1000, the aggregate norm ratio of insertion-time discrepancy to subsequent drift is 0.04 on CounterFact and 0.05 on ZsRE, with corresponding aggregate weight differences of 6% and 7% (Eq. (79), Appendix D.3.3). The final signal on this cohort is dominated by subsequent drift.

References

[1] Mass-Editing Memory in a Transformer.

[2] AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models.

We apologize for splitting this response due to the character limit and address the remaining W2 controls, W3, and W4 in Part 2.

AuthorsOpenReview ↗

Rebuttal by Authors (Part 2 of 2)

Part 1 explains how the anchor references differ. We continue W2 with the control experiments and then address W3 and W4.

W2(2) Anchor reference and the independent role of gradient drift (continued).

Table 7 (Appendix D.3.2) reports the requested controls after 1,000 sequential edits on LLaMA-3.1-8B-Instruct.

VariantCounterFact Hist. (%)ZsRE Hist. (%)
GDR90.0088.37
Post-update anchor87.0085.60
Query-only85.0084.50
Shuffled GDR weights76.5075.90

The post-update anchor preserves most of the gain. The original reference further improves retention. GDR exceeds query-only by 5.00 and 3.87 percentage points, respectively. Shuffling the same aggregate weight values reduces retention, supporting the importance of coordinate assignment.

Across these variants, Table 8 reports immediate-efficacy differences of at most 0.10/0.05 percentage points on CounterFact/ZsRE, with mean parameter-update norms within 1.8% of GDR.

Table 9 pairs writes from identical checkpoints and matches update norms per request. Mean historical displacement stays within 2% of GDR, and new-edit efficacy is comparable. GDR yields lower mean prediction damage on both datasets. On ZsRE, damage is 0.014 nats per target token for GDR, 0.024 for query-only, and 0.040 for shuffled weights. This supports coordinate-specific protection at matched update scales.

Table 10 (W1(3) in Part 1) also includes query-only. GDR achieves higher Hist. in every tested order, with means of 91.17% versus 87.50%. Sections 5.1 and 5.3 now link to these additional ablations and diagnostics.

W3 (Mathematical formulation).

We apologize for the imprecise formulation and thank the reviewer for identifying these issues. In Section 3, Eq. (4) now states a regularized least-squares objective, whose closed-form solution is Eq. (5), allowing a trade-off between target fitting and background preservation.

Appendix B.1 now defines V0=WoutK0V_0=W_{\mathrm{out}}K_0 relative to the current pre-write weights. Consequently, (V0−WoutK0)K0⊤=0(V_0-W_{\mathrm{out}}K_0)K_0^\top=0 holds exactly, justifying the cancellation when deriving Eq. (29) from Eq. (27). This reference need not be explicitly constructed. Appendix C.3 specifies the scaled background matrix C0=λCE^[kk⊤]C_0=\lambda_C\widehat{\mathbb E}[kk^\top] and its use in the update (Eq. (67)). We cache and reuse the second-moment estimates and define the output reference using the current weights.

We also corrected ZsRE specificity to use the locality question’s own reference answer (Eq. (66), Appendix C.1). Eq. (5) gives the same closed-form update under the clarified assumptions.

W4 (Scope of diagnostic interpretations).

(1) Fixed-cohort retention. Section 6.3 now explicitly clarifies that the earliest-20% cohort in Figs. 2 and 7 grows with TT. We retain this analysis and add Table 2, which tracks the same first 200 ZsRE edits on LLaMA-3.1-8B-Instruct throughout a shared 1,000-edit sequence. Drop is each method’s accuracy at T=200T=200 minus its accuracy at T=1000T=1000, in percentage points.

MethodT=200T=200T=500T=500T=1000T=1000Drop (pp) ↓
GDR97.2093.7288.378.83
AlphaEdit [2]96.9087.4576.2520.65
HSE [3]97.0589.1680.4216.63

Starting accuracies differ by at most 0.30 percentage points, while GDR loses 8.83 points over the subsequent 800 edits, compared with 20.65 and 16.63 for AlphaEdit and HSE. This directly tracks forgetting without changing cohort membership.

(2) Datasets and timing boundaries. Table 4 and Fig. 4 (Appendix D.2) use 1,000 CounterFact edits on LLaMA-3.1-8B-Instruct. Table 4’s caption now states this setup. Fig. 6 also uses CounterFact (Appendix D.6).

Appendix C.3 times each request from before processing through the write and state updates, including GDR anchor computation and history updates. CUDA synchronization occurs at both boundaries. Timing excludes model loading, one-time covariance preparation, evaluation, and checkpointing. Table 4 and Fig. 6(a) average all 1,000 steps. Fig. 6(b) times the next write at the stated history length (Appendix D.6).

(3) Edit Score [1]. Appendix D.2 now explicitly defines the metric used in Table 4 and Fig. 4 as Sedit=3/(Eff.−1+Gen.−1+Spe.−1)S_{\mathrm{edit}}=3/(\mathrm{Eff.}^{-1}+\mathrm{Gen.}^{-1}+\mathrm{Spe.}^{-1}), with all three components expressed on the same percentage scale. This makes the reported editing-quality criterion explicit.

References

[1] Mass-Editing Memory in a Transformer.

[2] AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models.

[3] Hippocampal-like Sequential Editing for Continual Knowledge Updates in Large Language Models.

We have carefully considered your feedback and provided responses and additional details where relevant, and we look forward to hearing your thoughts and continuing the discussion.

As you can see, weaknesses 1, 3, and 4 are minor issues; I pointed them out simply to facilitate your revisions. Regarding weakness 2, while I acknowledge the empirical correlation, the underlying theoretical issues remain difficult to resolve. Therefore, I would like to see the authors design more rigorous boundary experiments to substantiate this correlation. The current manuscript is insufficient to support this point; consequently, I am maintaining my original rating.

Reviewer_qm2qOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a knowledge editing method called GDR to improve performance under the sequential editing setting. The authors demonstrate that gradient drift in the editing residual space can govern changes in the model's local responses. Based on this finding, GDR saves gradient anchors of the current model when a new edit is inserted; during subsequent edits, it employs a Top-kk retrieval mechanism to identify affected historical samples and computes the discrepancy between the current gradients and the stored anchors, thereby constructing a regularization penalty to constrain model updates. Experiments show that GDR effectively improves the retention of previously edited knowledge while preserving the model's general capabilities.

Critical strengths
  • The theoretical analysis is rigorous and well-constructed. The perspective of analyzing knowledge editing through gradient drift is novel and provides clear theoretical grounding for the proposed method.
  • The proposed method achieves strong empirical results in large-scale sequential editing experiments, while also demonstrating good preservation of downstream task generalization ability.
Critical weaknesses
  • The proposed method still incurs overhead that grows with the number of historical edits. Specifically, it requires persistent storage of all historical editing samples' texts and their corresponding gradient anchors. This limits the long-term scalability of the approach and, to some extent, undermines its advantage as a parameter-modifying method (as opposed to memory-based approaches).
  • The method appears to be highly sensitive to hyperparameter settings. The regularization coefficient β\beta varies considerably across different model architectures. Moreover, different hyperparameter configurations seem to be required for different datasets; for instance, β\beta is set to 3.5×10−43.5 \times 10^{-4} for Qwen2.5 on ZsRE but 5×10−55 \times 10^{-5} on CounterFact. This suggests that the method heavily relies on extensive hyperparameter tuning, which raises concerns about its adaptability to real-world deployment scenarios where the data distribution may not be known in advance.
Questions for authors
  • How sensitive is GDR to the choice of β\beta? Could the authors provide a systematic hyperparameter sensitivity analysis (e.g., performance curves as β\beta varies) to clarify the robustness of the method across different settings?
1 comment
AuthorsOpenReview ↗

Rebuttal by Authors

Thank you very much for your careful review and constructive comments.

W1 (Historical storage and scalability).

We agree that retaining historical texts, keys, and gradient anchors introduces growing storage costs and limits long-term scalability.

To clarify their role, these records support subsequent editing, including historical selection and query-gradient computation. The edited knowledge itself is written into model parameters, and ordinary inference uses the edited model without retrieving these records. GDR therefore retains this inference property of parameter-modifying methods, while requiring auxiliary storage for continued editing.

Top-kk limits the expensive historical query-gradient evaluations to at most min⁡(20,N)\min(20,N) per step. Appendix D.6 reports the corresponding runtime decomposition: query-gradient time approximately plateaus over the measured history range, while key-selection time continues to grow.

For quantitative context, the following FP32 accounting gives the historical-vector storage for 1,000 records:

ModelKey + anchor (MiB)Including one cached residual per edit (MiB)
LLaMA-3.1-8B-Instruct70.3185.94
Qwen2.5-7B-Instruct85.9499.61
GPT-2-XL30.5236.62

These are calculated tensor payloads. Historical text/token data, metadata, covariance caches, and temporary workspaces require additional space and are excluded from this table.

The performance–time analysis in Appendix D.2 further supports the moderate budget: k=20k=20 achieves an Edit Score [1] of 83.4 at 20.18 seconds per edit, compared with 84.2 at 25.31 seconds for k=25k=25. This provides a useful trade-off in the evaluated setting. We have clarified the remaining storage and retrieval costs in Appendix D.6 and acknowledge their growth as a limitation in the conclusion.

W2 & Q1 (Sensitivity to the regularization coefficient β\beta).

Thank you for highlighting this concern. We have added a systematic sensitivity analysis in Appendix D.8 on two model architectures and both datasets, and clarified the reported configurations in Appendix C.3.

The effect of β\beta depends on the scale of LGDR\mathcal{L}_{\mathrm{GDR}} relative to the editing objective. Since GDR uses squared gradient discrepancies, scaling these discrepancies by cc scales the penalty by c2c^2. Numerical differences between coefficients therefore do not, by themselves, establish high sensitivity. We examine this directly through controlled coefficient sweeps.

Within each model–dataset configuration, we vary only β\beta, restarting from the same initial checkpoint and using the same 1,000-edit sequence. On LLaMA-3.1-8B-Instruct, the tested settings at 0.5×0.5\times, 1×1\times, and 2×2\times the original coefficient differ by at most 0.72 percentage points from the original results across all five reported metrics and both datasets. On Qwen, the tested coefficients {1,2,3.5}×10−4\{1,2,3.5\}\times10^{-4} keep all five metrics within 1 percentage point of their respective original configurations.

The analysis also identifies the expected trade-off: smaller coefficients weaken retention, while the largest tested coefficients reduce immediate efficacy or generalization.

Appendix C.3 records the coefficients used in the main experiments; these are not unique values required for good performance. For quantitative context, using a single β=2×10−4\beta=2\times10^{-4} on both Qwen datasets gives:

DatasetOriginal β\betaHist. (%) Original → SharedImmediate Eff. (%) Original → Shared
CounterFact5×10−55\times10^{-5}99.50 → 99.50100.00 → 99.70
ZsRE3.5×10−43.5\times10^{-4}90.42 → 90.1099.30 → 99.37

Across Eff., Gen., Spe., Hist., and immediate efficacy, every change is at most 0.32 percentage points in absolute value. This demonstrates that comparable performance does not require separate dataset-specific β\beta values in these tested configurations. The other optimization settings remain those documented in Appendix C.3.

These results clarify that GDR admits multiple effective coefficient choices while preserving a favorable acquisition–retention trade-off.

References

[1] Mass-Editing Memory in a Transformer.

We have carefully considered your feedback and provided responses and additional details where relevant, and we look forward to hearing your thoughts and continuing the discussion.

Reviewer_RLteOpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper studies catastrophic forgetting in sequential knowledge editing for LLMs. Authors establish a relationship between gradient discrepancies and changes in local responses, which motivates a new method called Gradient Drift Regularization (GDR). This method stores a gradient anchor when each edit is performed and subsequently compares it with the gradient computed under the current model. The resulting gradient discrepancy is used to construct a dimension-wise quadratic regularizer that penalizes new editing residuals along coordinates with larger historical gradient drift. To reduce computational overhead, they further introduce a selection mechanism that restricts gradient evaluations to a subset of historical edits. Experiments across multiple models demonstrate improvements in sequential editing performance and historical knowledge retention compared with existing methods.

Critical strengths
  1. The proposed regularization is relatively simple and can be integrated into an existing MEMIT-style editing framework without substantial architectural modifications.

  2. The experimental results are encouraging across multiple model families and knowledge editing benchmarks. The authors also provide useful ablation studies to examine the effectiveness of different components of the proposed method.

Critical weaknesses
  1. Theorem 1 establishes a bound on the difference between two centered local responses in terms of their gradient discrepancy. However, this result does not establish a direct connection between gradient drift and knowledge forgetting. Specifically, the bound does not constrain the difference between the predictions at the reference points, i.e., fA(0)f_A(0) and fB(0)f_B(0). Even when the gradient discrepancy is zero, the two models may assign substantially different probabilities to the historical target if fA(0)f_A(0) and fB(0)f_B(0) differ significantly. Therefore, the theoretical result does not explain why controlling gradient drift should prevent catastrophic forgetting.

  2. Even in the special case where fA(0)=fB(0)f_A(0)=f_B(0), the connection between the proposed regularizer and forgetting prevention remains unclear. Specifically, GDR penalizes new editing residuals along coordinates that have exhibited larger historical gradient drift. However, there is no evidence that these coordinates are more likely to be affected by future updates in ways that could damage historical edits. For example, a coordinate may exhibit negligible historical gradient drift while remaining highly sensitive to perturbations. Updating this coordinate could substantially alter historical predictions, yet GDR would impose almost no penalty. Thus, it is unclear why historical gradient drift, rather than the sensitivity of historical predictions to the current update, should serve as the coordinate-wise regularization weight.

  3. Although the top-k mechanism limits the number of query-gradient evaluations, GDR still stores a gradient anchor and a key representation for every historical edit. Consequently, its memory requirements grow linearly with the number of edits. Identifying the top-k historical edits also requires scoring all historical keys, which adds computational overhead that grows with the length of the editing history.

Questions for authors
  1. As discussed in Weaknesses 1 and 2, there appear to be two gaps in the motivation for GDR: the connection between gradient drift and knowledge forgetting, and the connection between the proposed drift-weighted regularizer and the reservation of edits. Could the authors provide a more principled explanation of why gradient drift is an appropriate signal for mitigating catastrophic forgetting and why the proposed regularization objective should be effective?

  2. In the current formulation, the anchor gradient and query gradient are evaluated at different reference centers. Specifically, the anchor gradient is computed at the optimized editing residual under the pre-edit model, whereas the query gradient is computed at zero residual under the current model. How do the authors interpret the effect of this discrepancy? Could evaluating gradients at different reference centers provide useful information, or might it introduce confounding effects into the measured gradient drift? More importantly, how might this design choice affect the effectiveness of GDR?

10 comments
AuthorsOpenReview ↗

Rebuttal by Authors (Part 1 of 2)

Thank you very much for your careful review and constructive comments.

W1 & Q1 (Connection between gradient drift and knowledge forgetting).

Equal gradients can coexist with different reference-point probabilities. Section 4 and Appendix B.3 clarify the role of gradient drift in the local response, and we evaluate its connection to forgetting empirically.

(1) What gradient drift characterizes. Let Δg=gA(ej)−gB(ej)\Delta g=g_A(e_j)-g_B(e_j) and C(ϵ)=[fA(ϵ)−fA(0)]−[fB(ϵ)−fB(0)]C(\epsilon)=[f_A(\epsilon)-f_A(0)]-[f_B(\epsilon)-f_B(0)]. Under Theorem 1's assumptions, Taylor expansion gives C(ϵ)=⟨Δg,ϵ⟩+R(ϵ)C(\epsilon)=\langle\Delta g,\epsilon\rangle+R(\epsilon) with ∣R(ϵ)∣≤L∥ϵ∥22|R(\epsilon)|\leq L\|\epsilon\|_2^2. The inner product gives the first-order difference in the two states' responses to the same local perturbation. The triangle inequality gives ∣fA(ϵ)−fB(ϵ)∣≤∣fA(0)−fB(0)∣+∥ϵ∥2∥Δg∥2+L∥ϵ∥22|f_A(\epsilon)-f_B(\epsilon)|\leq |f_A(0)-f_B(0)|+\|\epsilon\|_2\|\Delta g\|_2+L\|\epsilon\|_2^2 (Eq. (56)). The reference-score gap remains a separate term.

(2) Connection to observed forgetting. Appendix D.7.2 (Table 13) tracks the same 200 ZsRE edits on LLaMA-3.1-8B-Instruct at T∈{200,500,1000}T\in\{200,500,1000\}. We relate each edit's drift to Fj(T)=[ℓj(Gj)−ℓj(GT)]+F_j(T)=[\ell_j(G_j)-\ell_j(G_T)]_+, where ℓj\ell_j is target log-probability per token and GjG_j is the actual post-write state. The nine within-method and within-checkpoint Spearman correlations range from 0.6160.616 to 0.7910.791. At T=1000T=1000, the association remains positive after adjusting for insertion-time target confidence and answer length.

MethodAdjusted correlationMean cumulative loss Fˉ1000\bar{F}_{1000}
GDR0.6050.231
AlphaEdit [1]0.6670.579
HSE [2]0.6910.417

Loss is in nats per target token. Edits with larger drift tend to suffer greater loss within each method and checkpoint. GDR also has the lowest mean loss. On the fixed cohort in Table 2 (§6.3), GDR, AlphaEdit, and HSE have similar accuracy at T=200T=200 and lose 8.838.83, 20.6520.65, and 16.6316.63 percentage points, respectively, over the subsequent 800 edits. W2 tests how much protection the drift weights provide under controlled writes.

W2 & Q1 (Why drift-weighted coordinate regularization?).

A coordinate with a large, unchanged gradient can remain sensitive. GDR penalizes candidate displacements along coordinates whose local response has changed since insertion. Let Δgj=gquery(ej)−ganc(ej)\Delta g_j=g^{\mathrm{query}}(e_j)-g^{\mathrm{anc}}(e_j) and let Cj(ϵ)C_j(\epsilon) denote the difference between the centered query and anchor responses. For a small displacement along coordinate rr, Cj(ter)=tΔgj,r+O(t2)C_j(te_r)=t\Delta g_{j,r}+O(t^2), where ere_r is the coordinate unit vector. The squared first-order term is (tΔgj,r)2(t\Delta g_{j,r})^2. For a common candidate direction δ\delta, Cauchy–Schwarz gives ∑ej∈Bi⟨Δgj,δ⟩2≤d LGDR(i)(δ)\sum_{e_j\in\mathcal B_i}\langle\Delta g_j,\delta\rangle^2\leq d\,\mathcal L_{\mathrm{GDR}}^{(i)}(\delta), where d=dim⁡(δ)d=\dim(\delta) (Eq. (57), Appendix B.3). Section 5.3 uses this connection to motivate the diagonal penalty.

Table 7 (Appendix D.3.2) compares the weighting schemes on LLaMA-3.1-8B-Instruct after 1,000 sequential edits.

VariantCounterFact Hist. ↑ZsRE Hist. ↑
GDR90.0088.37
Query-only85.0084.50
Shuffled GDR weights76.5075.90

GDR improves historical efficacy over query-only by 5.00 and 3.87 percentage points. Shuffling preserves the aggregate weight values and changes their coordinate assignment, reducing historical efficacy by 13.50 and 12.47 points.

Table 9 (Appendix D.3.3) reports paired single writes from identical checkpoints with parameter-update norms matched for each request. The following results are on ZsRE.

VariantNew Eff. (%)Mean relative historical displacementDˉ↓\bar{D}\downarrow
GDR99.101.0000.014
Query-only99.081.0100.024
Shuffled GDR weights99.001.0200.040

Dˉ\bar{D} measures prediction damage from one write in nats per target token. The variants have similar immediate efficacy and mean historical displacements within 2% of GDR. GDR's lower mean damage under these controls supports using the historical reference and assigning the drift weights to their corresponding coordinates.

References

[1] AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models.

[2] Hippocampal-like Sequential Editing for Continual Knowledge Updates in Large Language Models.

We apologize for splitting our reply due to the character limit and address Q2 and W3 in Part 2.

AuthorsOpenReview ↗

Rebuttal by Authors (Part 2 of 2)

Following Part 1 on the role and effects of gradient drift, we address Q2 on the reference centers and W3 on storage and retrieval cost.

Q2 (What does the measured gradient difference contain?).

The two gradients deliberately use different references. The anchor records the response around the optimized insertion-time representation, while the query measures the response around the current representation. Both differentiate with respect to the same additive edit-space coordinates. Theorem 1 explicitly accommodates these state-specific reference centers. Their practical influence is quantified and tested, rather than assumed negligible.

To distinguish insertion-time reference effects from subsequent changes, we compute gpost(ej)g^{\mathrm{post}}(e_j) immediately after the actual parameter write on GjG_j, with zero additional residual. Appendix D.3.3 then decomposes the signed gradient difference into bj=gpost(ej)−ganc(ej)b_j=g^{\mathrm{post}}(e_j)-g^{\mathrm{anc}}(e_j) and aj(t)=gtquery(ej)−gpost(ej)a_j^{(t)}=g_t^{\mathrm{query}}(e_j)-g^{\mathrm{post}}(e_j), giving gtquery(ej)−ganc(ej)=aj(t)+bjg_t^{\mathrm{query}}(e_j)-g^{\mathrm{anc}}(e_j)=a_j^{(t)}+b_j.

Here, bjb_j captures the difference between the optimized-target reference and the realized write, including the associated model and representation changes. The term aj(t)a_j^{(t)} captures changes after the actual write. Thus, we explicitly account for the contribution identified by the reviewer.

On the original GDR trajectory at T=1000T=1000, we evaluate the same first 200 edits. The reference/later-change ratio is ∑j∥bj∥22/∑j∥aj(1000)∥22\sqrt{\sum_j\|b_j\|_2^2/\sum_j\|a_j^{(1000)}\|_2^2}. We also measure the relative aggregate-weight difference when replacing the original anchor with the post-update anchor. Appendix D.3.2 separately reports the corresponding full-sequence retention results.

DatasetReference/later-change ratioAggregate-weight differenceGDR Hist.Post-anchor Hist.
CounterFact0.046%90.0087.00
ZsRE0.057%88.3785.60

The weight difference is measured on the same GDR trajectory using relative ℓ1\ell_1 distance, with post-anchor weights as the denominator. These final-cohort measurements show that the insertion-time discrepancy does not dominate the aggregate drift signal, while still allowing it to affect individual coordinate weights.

Changing to a post-update reference retains much of GDR's performance. Nevertheless, the original anchor improves historical efficacy by 3.00 and 2.77 percentage points at comparable immediate edit efficacy. This supports retaining the optimized-target response as the fixed reference in the evaluated setting. It does not require the insertion-time discrepancy to be zero or negligible at every editing step.

We have clarified the reference meanings in Section 5.1 and provided the decomposition, quantitative diagnostics, and reference-replacement controls in Appendices D.3.2–D.3.3. For further details on the reference controls, please also refer to our response to W2 for Reviewer qBDE.

W3 (Historical storage and retrieval cost).

Top-kk limits the expensive historical query-gradient evaluations to at most min⁡(20,N)\min(20,N) per step. We agree that historical-record storage and exact key scoring still grow with NN. This remains a limitation of the current implementation.

In Fig. 6(b), query-gradient runtime approximately plateaus once the selection budget is reached, while key-selection time continues to grow over the measured history range. Thus, the benefit is a bounded historical gradient workload, with the remaining retrieval cost explicitly accounted for.

For storage, the following FP32 accounting quantifies the historical-vector payload for 1,000 edits.

ModelKey + anchor (MiB)Including one cached residual per edit (MiB)
LLaMA-3.1-8B-Instruct70.3185.94
Qwen2.5-7B-Instruct85.9499.61
GPT-2-XL30.5236.62

These are calculated tensor payloads, excluding text/token data, metadata, covariance caches, and temporary workspaces. They quantify the storage required by the historical vectors without treating it as total memory.

The measured performance–time trade-off in Appendix D.2 further supports the moderate budget of k=20k=20 on LLaMA-3.1-8B-Instruct and CounterFact.

kkEdit Score [3] ↑Time/Edit (s) ↓
2083.420.18
2584.225.31

The k=20k=20 setting requires 20.3% less time than k=25k=25, with a score difference of only 0.8 points. These results support the selected performance–cost balance. We have clarified in Appendix D.6 that top-kk controls historical gradient computation, while storage and full-history retrieval continue to grow.

References

[3] Kevin Meng et al. (ICLR 2023). Mass-Editing Memory in a Transformer.

We have carefully considered your feedback and provided responses and additional details where relevant, and we look forward to hearing your thoughts and continuing the discussion.

Thank you for the detailed response and the additional experiments. My concerns in Q2 and W3 have been addressed. However, I believe the two central concerns (W1 and W2) remain unresolved.

  1. Regarding W1, the newly added Eq. (56) reinforces my concern:

∣fA(ϵ)−fB(ϵ)∣≤∣fA(0)−fB(0)∣+∥ϵ∥∥Δg∥+L∥ϵ∥2.|f_A(\epsilon)-f_B(\epsilon)| \le |f_A(0)-f_B(0)| + \Vert\epsilon\Vert \Vert\Delta g\Vert + L\Vert\epsilon\Vert^2.

Controlling the second term places no constraint on the first, so it does not guarantee that historical predictions are preserved.

The new experiments show that gradient drift is positively correlated with forgetting, which I do not dispute. My question is why the first term can be ignored, so that controlling only the second term mitigates catastrophic forgetting. The goal of sequential editing is precisely to prevent many subsequent edits from substantially changing historical predictions. If the authors believe the first term is negligible or is controlled indirectly, I would expect a theoretical argument or a targeted experiment.

  1. Regarding W2, the response does not answer my question: why are coordinates with larger historical gradient drift more likely to be affected by future edits, and therefore in need of a stronger penalty?

Under the MEMIT update with a single layer and fixed keys, a new edit ii changes the value of historical edit jj by

Δvj=ΔWkj=αijri,αij=ki⊤(K0K0⊤+kiki⊤)−1kj.\Delta v_j = \Delta W k_j = \alpha_{ij} r_i, \quad \alpha_{ij} = k_i^\top (K_0K_0^\top + k_ik_i^\top)^{-1} k_j.

This change is determined by the key interaction αij\alpha_{ij} and the current residual rir_i. It remains unclear why coordinates with larger historical drift are more susceptible to interference from future edits, or how penalizing them reduces that interference more effectively.

The new Eq. (57) bounds the difference between two centered local responses, not the effect of the actual parameter update on historical knowledge, so it does not directly address this question.

Overall, I appreciate the additional experiments, and I agree that the new ablations provide empirical support for the method. However, my concerns on W1 and W2 remain, and I hope the authors can clarify them, including the mechanism behind the regularizer beyond its empirical effect.

Comments on the paper

1 comment

Dear Area Chair,

Thank you for your time and consideration. Following the reviewers’ feedback, we have revised the manuscript and added targeted experiments to strengthen its empirical support and clarify the scope of its theoretical claims.

  • Following reviewer RLte’s comments, we added a shared-checkpoint continuation experiment to directly measure historical prediction deterioration over successive writes. After 200 further edits, the mean historical score drop was 0.103 for GDR, compared with 0.152 for query-only and 0.271 without historical regularization, with comparable immediate editing efficacy. We also quantified anchor and drift contributions along actual write-induced displacements, while explicitly distinguishing these empirical findings from a general preservation guarantee.

  • Following reviewer qBDE’s comments, we added matched-update comparisons with query-only, post-anchor, and shuffled-weight controls, together with a boundary experiment grouping requests by their pre-write weight differences. These analyses examine when the original anchor provides additional protection and when its contribution is small. We also revised the statistical reporting, mathematical presentation, and fixed-cohort retention analysis.

  • Following reviewer qm2q’s comments, we added systematic regularization-coefficient sensitivity experiments across two model architectures and both datasets. A shared coefficient across the two Qwen datasets changed the reported metrics by at most 0.32 percentage points. We also clarified the remaining historical storage and retrieval costs.

These additions improve the manuscript’s experimental coverage, methodological clarity, and discussion of limitations. We appreciate the reviewers’ constructive feedback and your consideration of the revised manuscript and subsequent discussions.

Best wishes,
The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.