Edit Less, Achieve More: Dynamic Sparse Neuron Masking for Lifelong Knowledge Editing in LLMs

W1 & Q1: Larger batch sizes:

We thank the reviewer for suggesting evaluating larger batch sizes. Following your advice, we conducted experiments with batch sizes 4 and 10 on 2000 sequential edits to better simulate practical simultaneous update scenarios. Due to rebuttal time constraints, we only present results for batch size=4 below. We will report the full editing result as soon as we finish the experiments.

Notably, we optimized AlphaEdit’s evaluation by evaluating only after all edits; this revised approach is fairer, more reasonable, and more challenging. Our findings show that NMKE maintains stable editing accuracy and locality as the number of edits increases, while AlphaEdit’s performance degrades significantly. This is because AlphaEdit uses a fixed global projection matrix computed once, which cannot adapt to distributional shifts over multiple edits. In contrast, NMKE employs fine-grained neuron-level masking, which limits weight modifications per batch and mitigates model drift.

References:

[1] LEMoE: Advanced Mixture of Experts Adaptor for Lifelong Model Editing of Large Language Models. EMNLP'24

Revision. We will include the complete experimental results for batch sizes 4 and 10 in the manuscript and clarify the practical advantages of NMKE under both sequential and batch editing scenarios.

W2 & Q2: Efficiency and runtime analysis.

Thank you for your thoughtful questions. We provide a detailed analysis from two perspectives: memory overhead and per-edit computational cost.

Memory Overhead:

To minimize overhead, our method computes neuron attribution only for the edited layers (4–8), which is consistent with AlphaEdit. For each layer, the following additional memory objects are instantiated:

• Attribution coefficients (~0.33 MB per layer)
• Binary importance mask (~0.05 MB per layer)
• Two scalar values (resonance and burst ratios)

The statistics from layers 4 to 8 are summarized in the table below:

The intermediate tensors involved account for 0.006248% of the LLaMA-3-8B model's memory. For our method, the additional memory consumption is very marginal.

Computational Overhead:

To address your concerns about the computations introduced by attribution, we report the per-edit runtime and memory usage of current SOTA methods and NMKE with different attribution strategies (MLP Projection Coefficient (MPC) [R2], probability Shift Attribution(PSA) [R1], and Log Probability Shift (LPS) [R1]) under sequential edits (after $T$=1,000 and $T$=2,000 edits).

1. The most efficient variant NMKE (MPC) outperforms AlphaEdit across all metrics at $T$=2,000, with only ~1s extra runtime and minimal memory growth.
2. The most advanced variant NMKE (LPS) remains practically efficient, adding only 8s per edit compared to AlphaEdit. Over 2000 edits, this adds ~4.5 hours, an acceptable cost for substantial quality gains: Locality rises from 0.14 to 0.74, with Reliability and Generalization each up 0.32.
3. NMKE is a flexible framework that supports attribution methods of varying complexity, enabling users to balance quality and efficiency under different scenarios.

References:

[1] Neuron-level knowledge attribution in large language models. arXiv'23.

[2] Transformer Feed‑Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Space, EMNLP'22.

Revision. We will add a new subsection in Appendix C that details all additional memory objects and their sizes for LLaMA3‑8B, and includes a quantitative table clearly comparing NMKE’s computational overhead with other methods.

L1: Cross-layer interactions in LLMs.

Thank you for your insightful and constructive suggestion. Although NMKE performs neuron attribution for each layer separately, the attribution module is integrated into the end-to-end optimization framework, which may help implicitly learn cross-layer neuron interactions. However, we agree with your suggestions that explicitly modeling cross-layer interactions like CLT can result in more accurate parameter localization and better editing performance. We will explore this as a key direction in our future work.

Revision. We will expand the discussion of cross-layer knowledge tracing methods in the limitations section, and outline how integrating such perspectives could enhance NMKE’s attribution and masking in future work.

W1: Binary classification overlooks the continuity of neuron functions. & Q4: soft masking approach:

We sincerely thank the reviewer for this profound insight. To evaluate the sufficiency of distinguishing different functions of neurons and omitting subsets during editing, we replaced the binary-masking strategy with the soft-masking strategy to incorporate the continuity of the neuron states. Specifically, the update magnitude for each neuron is directly scaled by its attribution score.

We conducted ablation experiments on 2000 sequential edits. The comparison between our default NMKE and the NMKE-softmask variant is shown below:

• Soft masking matches NMKE in accuracy and generalization but reduces locality, due to the editing of all neurons.
• NMKE avoids this by selectively editing subsets of the neurons, maintaining stability and locality in lifelong editing.

Revision. We will explore scaling neuron update magnitudes (e.g., learning rates) by continuous attribution scores for more nuanced, less disruptive edits in future work.

W2: computational cost analysis

Thank you for the insightful suggestion. To address your concerns of the computations introduced by attribution, we report the per-edit runtime and memory usage of current SOTA methods and NMKE with different attribution strategies (MLP Projection Coefficient (MPC) [R2], probability Shift Attribution(PSA) [R1], and Log Probability Shift (LPS) [R1]) under sequential edits (after $T$=1,000 and $T$=2,000 edits).

1. The most efficient variant NMKE (MPC) outperforms AlphaEdit across all metrics at $T$=2,000, with only ~1s extra runtime and minimal memory growth.
2. The most advanced variant NMKE (LPS) remains practically efficient, adding only 8s per edit compared to AlphaEdit. Over 2000 edits, this adds ~4.5 hours, an acceptable cost for substantial quality gains: Locality rises from 0.14 to 0.74, with Reliability and Generalization each up 0.32.
3. NMKE is a flexible framework that supports for attribution methods of varying complexity, enabling users to balance quality and efficiency under different scenarios.

References:

[R1] Neuron-level knowledge attribution in large language models. arXiv'23.

[R2] Transformer Feed‑Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Space, EMNLP'22.

Revision. We will add a computational cost analysis and comparison with other methods in the Appendix.

W3: Unclear and non-generalizable hyperparameter selection in dynamic masking. & Q2: Selection and generalizability of entropy-based hyperparameters.

We would like to clarify that parameters such as the temperature coefficient $a$ and scaling factors $a_{ge}$/$a_{sp}$ were not manually tuned, but set based on observed neuron activation patterns. This approach enables the mask to adapt to each batch’s activation patterns, typically selecting 30–40% of neurons without manual tuning. The masking ratio is derived directly from entropy, and parameters like $a_{ge}$ and $a_{sp}$ simply scale the outputs rather than requiring task-specific adjustment.

We used the same hyperparameters across all models and datasets, and NMKE consistently achieves strong editing and generalization performance. The sensitivity analysis (Figure 7a) further confirms NMKE’s robustness to a wide range of hyperparameter values.

Revision. We will clarify the hyperparameter selection and supplement the appendix with extra experiments to further validate the robustness and generalization.

W4: More LLMs like Qwen

Thank you for the suggestion. We have implemented our method on Qwen2.5-7B and performed 2,000 edits. Due to time constraints, results from the first 100 edits are below. We will report the full editing result as soon as we finish the experiments.

The results show our method performs similarly in the first 100 edits on Qwen2.5 and LLaMA3. This further demonstrates that our approach is model-agnostic and can be applied to any LLM with standard MLP layers.

Revision. We will provide the full Qwen results and a discussion of model compatibility in the revision.

Q1: Quantitative analysis of parameter stability

Thank you for the suggestion. We quantified weight shifts in layers 7–9 using Wasserstein, cosine, and L2 distances.

NMKE produces much lower distributional shifts than AlphaEdit in edited layers (4–8), better preserving internal weight structure. In the unedited layer (9), both methods show no change.

Revision. We will add quantitative ablation experiments on parameter shift to complement the t-SNE visualizations.

Q3: How do overlapping neurons influence the trade-off between editing success and locality?

We thank the reviewer for their insightful comments. To evaluate overlapping neurons, we edited using only overlapping or only non-overlapping neurons. The results are as follows:

Our ablation study reveals consistent patterns:

• Editing only overlapping neurons achieves the highest locality (0.80) while maintaining a certain level of accuracy (0.75).
• Editing only non-overlapping neurons, though it improves accuracy, reduces locality.
• By incorporating overlapping neurons, NMKE strikes the best balance between edit accuracy and locality.

Revision. We will add experiments (such as spectral analysis and neuron continuity visualization) and further clarify entropy-guided ratios and overlapping neurons, and their impact on the editing-locality trade-off.

W1: Discretely classify neurons into "general/specific" categories without considering the continuous spectrum of importance

Thank you very much for your suggestions. To evaluate the sufficiency of distinguishing different functions of neurons and omitting subsets during editing, we replaced the binary-masking strategy with the soft-masking strategy to incorporate the continuity of the neuron states. Specifically, the update magnitude for each neuron is directly scaled by its attribution score.

We conducted ablation experiments on 2000 sequential edits. The comparison between our default NMKE and the NMKE-softmask variant is shown below:

• Soft masking matches NMKE in accuracy and generalization but reduces locality, due to the editing of all neurons.
• NMKE avoids this by selectively editing subsets of the neurons, maintaining stability and locality in lifelong editing.

Revision. We will add ablation experiments and discuss the advantages of employing the binary-masking strategy compared to the soft-masking strategy.

W2: Relies on existing methods for neuron attribution approaches [1].

Thank you for the thoughtful questions. While our neuron attribution strategy is inspired by the perturbation-based analysis in [1], we would like to clarify that the core contribution of our paper, rather than proposing a new attribution technique, lies in building a flexible, modularized framework for knowledge editing that leverages neuron-level importance signals to enable fine-grained, dynamic, and selective edits.

Importantly, NMKE is not tied to a single attribution method and remains compatible with various strategies, including:

• MLP Projection Coefficient (MPC) [R2]: estimates importance via MLP down-projection weights, providing fast and low-overhead attribution.
• Probability Shift Attribution (PSA): computes the change in output probabilities under neuron ablation.
• Log-Probability Attribution (LPS) [R1]: measures the change in log-probability of the target token after neuron perturbation, offering the most faithful estimate of neuron contribution.

To demonstrate this flexibility, we conducted ablation studies using all three attribution methods within NMKE, showing that editing effectiveness is robust across variants:

These results show that NMKE delivers robust editing performance across attribution methods, consistently maintaining high accuracy, generalization, and locality.

References:

[R1] Neuron-level knowledge attribution in large language models. arXiv'23.

[R2] Transformer Feed‑Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Space, EMNLP'22.

Revision. We will add ablation results for multiple neuron attribution methods in the experiments section to demonstrate NMKE’s flexibility, and will briefly highlight its core contributions of dynamic masking, neuron categorization, and a flexible modularized framework.

W3: Requires manual hyperparameter tuning, impacting usability and efficiency.

We would like to clarify that parameters such as the temperature coefficient $a$ and scaling factors $a_{ge}$/$a_{sp}$ were not manually tuned, but rather set according to the distributional characteristics of neuron activations. Knowledge-general neurons typically show distributed and broad-domain activations, while knowledge-specific neurons are more focused on fine-grained facts.

Except for these parameters, we used the same hyperparameters across all models and datasets, with NMKE consistently achieving strong editing and generalization performance. Sensitivity analysis (Figure 7a) further confirms NMKE’s robustness to a wide range of hyperparameter values.

Revision. We will clarify the hyperparameter selection process in the main text, and supplement the appendix with additional experiments (e.g., hyperparameter sensitivity tests across more model-task combinations) to further validate robustness and generalization.

W4: There are a number of related works that have not been discussed and compared.

Thank you for highlighting these important related works. We appreciate the opportunity to clarify their distinctions and will incorporate a discussion of these methods in the revision.

• Ni et al. (F-Learning) proposed a two-stage parametric approach (forgetting followed by learning) for model-wide parameter updates, while NMKE uses neuron-level dynamic sparse masking to achieve precise, localized edits and reduce parameter drift.
• Pan et al. (FiNE) preserve language capabilities by freezing later layers, while NMKE constructs masks by identifying knowledge neurons, enabling fine-grained edits that preserve the model's general capabilities.
• Bi et al. (ATBias) proposed an in-context editing technique that biases key entity tokens during decoding without altering parameters, relying on external memory. In contrast, NMKE internally edits relevant neurons, offering an efficient large-scale knowledge editing approach without the need for external memory.

Revision. We will expand our discussion of the aforementioned related works to detail the distinctions and connections between these approaches and NMKE.

W1: Some gaps in description and complexity analysis.

Thank you for the careful review. We have revised the paper for precise terminology and tightened the exposition to avoid confusion and improve readability. Concrete clarifications with quantitative memory/runtime complexity are provided in Q1–Q3.

Revision. We will clarify technical terms and add a quantitative summary of NMKE’s memory and runtime complexity in Section 3 and Appendix C.

W2 & Q3: Illegible figures.

Thank you for your constructive feedback. We have revised all the visualizations by enlarging legend/axis fonts, increasing the line widths, and enhancing the color contrast; printed copies confirm these updates have greatly improved readability.

Revision. We will update Figures 2, 5, and 7 with enlarged fonts and higher-resolution vector graphics, and ensure all figures are clear in both print and digital versions.

Q1: Memory overhead and computational cost analysis.

We thank the reviewer for the thoughtful questions. Below, we provide a detailed analysis from memory overhead and per-edit computational cost.

Memory Overhead:

To minimize overhead, our method computes neuron attribution only for the edited layers (4–8), which is consistent with AlphaEdit. For each layer, the following additional memory objects are instantiated:

• Attribution coefficients (~0.33 MB per layer)
• Binary importance mask (~0.05 MB per layer)
• Two scalar values (resonance and burst ratios)

The statistics from layers 4 to 8 are summarized in the table below:

The intermediate tensors involved account for 0.006248% of the LLaMA-3-8B model's memory. For our method, the additional memory consumption is very marginal.

Computational Overhead:

To address your concerns about the computations introduced by attribution, we report the per-edit runtime and memory usage of current SOTA methods and NMKE with different attribution strategies: MLP Projection Coefficient (MPC) [R2], probability Shift Attribution(PSA) [R1], and Log Probability Shift (LPS) [R1] under sequential edits (after T=1,000 and T=2,000 edits).

1. The most efficient variant NMKE (MPC) outperforms AlphaEdit across all metrics at $T$=2,000, with only ~1s extra runtime and minimal memory growth.
2. The most advanced variant NMKE (LPS) remains practically efficient, adding only 8s per edit compared to AlphaEdit. Over 2000 edits, this adds ~4.5 hours, an acceptable cost for substantial quality gains: Locality rises from 0.14 to 0.74, with Reliability and Generalization each up 0.32.
3. NMKE is a flexible framework that supports attribution methods of varying complexity, enabling users to balance quality and efficiency under different scenarios.

References:

[R1] Neuron-level knowledge attribution in large language models. AarXiv'23.

[R2] Transformer Feed‑Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Space, EMNLP'22.

Revision. We will add a new subsection in Appendix C that details all additional memory objects and their sizes for LLaMA3‑8B, and includes a quantitative table clearly comparing NMKE’s computational overhead with other methods.

Q2: On the two editing tasks, which data partitions were used to estimate I and the scores (4-7) and when? On I. 173 you mentioned the ratio $\rho_{ge}$ is being "predicted" which confused me - I suspect to have missed something. What is the predictor here?

Thank you for the thoughtful question and careful reading.

1. For the two editing tasks, the neuron attribution matrix $I^{(l)}$ and scores (Eqs. 4-7) are estimated using the editing prompts from the current batch of requests, with no separate data partitions; Specifically, at each editing step, we use the current batch of editing requests as input prompts.

2. Regarding the confusion around "predicted" for $\rho_{ge}$ on line 173, we appreciate for pointing out the imprecise wording. There is no learned predictor; instead, $\rho_{ge}$ is estimated deterministically. We compute the average entropy $H_{ge}$ across neuron attribution distributions over the batch of prompts (Eq. 6), then linearly map this entropy to a selection ratio using fixed scalers $a_{ge}$ and bias terms $b_{ge}$ (i.e., $\rho_{ge} = H_{ge} \cdot a_{ge} + b_{ge}$), with no trainable parameters involved.

Revisions. We will replace "predicted" with "estimated" throughout (notably line 173) and clarify in Section 3.2 that ratios and scores are computed from the current batch of editing prompts without learned predictors.

L1: Lack of explicit discussion on limitations, such as not evaluating the effect of edits on related but distinct facts.

We truly appreciate your insightful observation. Your suggestion touches on a critical yet underexplored aspect of model editing: how edits to specific factual triples propagate to closely related but distinct knowledge, such as aggregated statistics or temporal sequences.

• Analysis method: Following your suggestion, we designed a pilot study to evaluate this phenomenon. We constructed 10 synthetic samples to simulate your highlighted scenario: updating a company’s CEO and querying related info like total CEO count or previous holders.
• Results: Notably, after updating a key fact (e.g., “Tim Cook → Elon Musk”), LLMs often failed to adjust related dependent facts; for example, they incorrectly estimated the company’s total CEO count, revealing post-edit inconsistency and factual conflicts.
• Future direction: This remains an important open yet challenge in knowledge editing. We plan to explore it further by constructing a dedicated benchmark for factual dependencies and incorporating metrics to assess post-edit consistency of related facts.

Revision. We will add a limitation discussion in the main paper regarding the lack of evaluation on the effects of edits on closely related facts, and outline this as a key direction for future work.

This is a gentle reminder that the authors have responded to your question. Could you please provide your feedback on their response?

Kindly note that the deadline of August 6 is approaching, and your timely feedback would be greatly appreciated to facilitate any further discussions if needed.

Thank you sincerely for your careful review and constructive suggestions, which have significantly enhanced the quality of our manuscript. We note your insightful points on the knowledge editing domain mentioned in the Limitations, and we will incorporate these into our future work. We will further refine the manuscript to ensure it meets the standards for acceptance.