acceptodds
Under review as a conference paper at ICLR 2027

Factual Association Editing without Pretrained Knowledge Representations

Abstract

We propose a closed-form factual association editing method for language models that completely eliminates the need to precompute and preserve pretrained knowledge representations. Unlike existing approaches that rely on proxy representations of pretrained knowledge obtained from sampled datasets, the proposed method formulates factual association editing as a well-conditioned optimization problem that enables stable and efficient editing using only new knowledge. During the editing process, we apply a principled treatment of ill-conditioning via within-representation centering and truncated singular value decomposition to prevent unstable or excessively amplified updates. The experimental results show that the proposed approach achieves editing performance comparable to that of state-of-the-art methods that rely on pretrained knowledge representations, while reducing amortized runtime for editing by up to 30, i.e., from 188 to 6 minutes, and requiring zero storage for pretrained knowledge representations. Even in large-scale editing (e.g., 2,000 edits), the proposed method maintains competitive editing performance, exhibiting only a 0.56% degradation on GPT-J without using any pretrained knowledge representations at all. These results provide empirical evidence that competitive factual association editing can be achieved without constructing an external proxy for the knowledge to be preserved. The code is available on an [anonymous repository](https://anonymous.4open.science/r/Anonymous-FCB2).

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Knowledge Editing, Large Language Models, Factual Association Editing, Well-Conditioned Optimization, Scalable Model Editing

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_dc1VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates an improved method for knowledge editing in LLMs. Existing approaches (e.g., MEMIT, AlphaEdit) typically rely on a pre-computed and stored hidden representation matrix K0K_0 of pre-training knowledge to constrain the editing process and preserve retain knowledge. The primary objective of this paper is to completely eliminate K0K_0 and construct the optimization problem solely based on the representation of new knowledge, K1K_1. The authors subsequently propose a series of techniques to address the challenges stemming from the removal of K0K_0. Experimental results demonstrate that the proposed method achieves promising performance while eliminating the pre-computation and storage overhead associated with K0K_0.

Critical strengths
  • The proposed method effectively eliminates the time overhead of pre-computation and reduces the amortized cost per edit at the experimental scale.
  • The method demonstrates solid empirical performance, achieving competitive results compared to several baselines designed for sequential editing on large-scale tasks.
Critical weaknesses
  • My primary concern regarding this paper lies in its core objective: whether removing the pre-computation of K0K_0 holds practical significance in real-world scenarios. In practice, computing K0K_0 does not appear to be prohibitively expensive, as it only needs to be computed once per LLM model, after which subsequent editing operations can be executed infinitely using the same K0K_0. Compared to ongoing edits, a few hours of one-time pre-computation cost is relatively low, and its GPU memory overhead during computation is also lower than that of various editing methods.
  • The proposed method still exhibits limitations in certain scenarios. For instance, in Table 3 (multi-hop reasoning evaluation), when NN exceeds 2000, the model's reasoning accuracy shows a noticeable decline compared to traditional methods. This suggests that there remains a performance gap in retaining reasoning capabilities when compared to traditional methods relying on pre-computed K0K_0.
Questions for authors
  • Could the authors provide further clarification on the targeted application scenarios? Specifically, in what practical settings does eliminating the pre-computation of K0K_0 offer a clear, decisive advantage?
  • Can the proposed framework be adapted as a general sequential editing method (e.g., by retaining K0K_0 while incorporating the newly introduced constraints to protect unrelated knowledge)? If so, how does its performance compare against other sequential editing baselines under this setup?
1 comment

Thank you for your thoughtful feedback and for recognizing our method's competitive performance. Below, we respond to each of your concerns in turn.

[Weakness 1 & Question 1. Objective]

We agree that the one-time precomputation cost can be amortized across repeated editing sessions. Our goal is to enable factual association editing without relying on a reference corpus to represent the model's existing knowledge. Moir [1] provides further motivation for examining this dependency. It shows that covariance estimates derived from static reference corpora may not adequately reflect the model’s distribution after post-training. Its findings highlight the importance of selecting reference data that adequately represent the knowledge to be preserved. Consequently, covariance estimates based on K0K_0 may be biased toward representation directions that are prominent in the reference corpus, potentially limiting the preservation of knowledge that is poorly represented in that corpus.

Our method eliminates the need to select such reference data or to construct a corresponding preservation covariance. This is particularly relevant when editing task-adapted models for which a suitable reference corpus representing their existing knowledge and capabilities is difficult to identify or unavailable. We have clarified this scenario in Section 5 in the revised manuscript.

[Weakness 2. Limitation]

We acknowledge the performance gap in multi-hop reasoning when the number of edits exceeds 2,000, as reported in Table 3. In the revised manuscript, the same results previously presented in Table 3 are shown in Figure 6 for easier comparison. We have explicitly discussed this limitation in Appendix A in the revised manuscript. The experiments below address the separate question of whether our framework can be extended to sequential editing. When we applied the extension proposed in the response to Question 2, accuracy decreased from 7.58 at N=1,000N=1,000 to 4.25 at N=2,000N=2,000. The drop in accuracy was smaller than before.

[Question 2. Sequential editing]

Yes, our framework can be adapted to sequential editing. Following your suggestion, we compared two variants of our method with existing sequential editing baselines. We conducted experiments on ZSRE using GPT-2 XL with a batch size of 100 for N=1,000N=1,000.

We evaluated a variant denoted as Ours (w/ K0K_0) that retains K0K_0 for knowledge preservation while applying centering and TSVD to K1K_1. We also developed a variant denoted as Ours** (w/o K0K_0) that supports sequential editing without requiring K0K_0. Specifically, we applied ridge regularization to penalize the update magnitude in all directions and incorporated statistics from previous edits to reduce interference with earlier edits.

MethodEfficacyGeneralizationSpecificity
MEMIT (w/ K0K_0)77.772.025.9
AlphaEdit (w/ K0K_0)92.986.825.6
Ours (w/ K0K_0)77.671.726.0
Ours** (w/o K0K_0)87.681.525.3

We also compared Ours** with UltraEdit on ZSRE using Mistral-7B with a batch size of 10 for N=1,000N=1,000.

MethodEfficacyGeneralizationSpecificity
UltraEdit79.572.947.9
Ours**97.291.555.4

Overall, these experiments support the feasibility of extending the proposed framework to sequential editing without K0K_0.

Thank you again for your constructive feedback. We have included these results in Appendix B in the revised manuscript.

[1] Kwon, J., Kim, J., Kim, D.K. and Cha, M., 2026. Moir: Let the model direct its own story for robust cross-domain knowledge editing. In Advances in Neural Information Processing Systems.

Reviewer_K2kiOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a closed-form method for factual association editing that eliminates the need to precompute and store representations of existing knowledge. Building on the linear associative-memory formulation of feed-forward layers, the method removes the explicit knowledge-preservation term and computes weight updates using only the new editing representations and target residuals. To address the resulting ill-conditioning, it combines direct pseudoinverse computation, within-representation centering, and truncated SVD, with the retained rank determined by spectral entropy or a cumulative fourth-power singular-value criterion. Experiments on GPT-2 XL, GPT-J, and Llama 3 demonstrate competitive performance on batch factual editing benchmarks while eliminating precomputation costs and reducing amortized editing time. Additional evaluations reveal limitations in sequential editing, multi-hop reasoning, and preservation of general capabilities.

Critical strengths
  1. I enjoy reading this paper. This paper is well-motivated and well-structured.
  2. The topic is timing and may provide insights for future works in the community.
Critical weaknesses

The method removes the explicit constraint on existing knowledge, but does not clearly explain why centering and truncated SVD can compensate for it. Let R=V1−WK1R=V_1-WK_1 and CK1=UΣQ⊤CK_1=U\Sigma Q^\top. For fixed layer inputs, the change in the outputs associated with existing knowledge is

ΔK0=RQqΣq−1Uq⊤K0.\Delta K_0=RQ_q\Sigma_q^{-1}U_q^\top K_0.

This depends on the overlap Uq⊤K0U_q^\top K_0 between the retained editing subspace and existing knowledge representations. Truncation controls amplification through the inverse singular values, but does not account for this overlap. Similarly, it is unclear why removing the all-ones direction through centering should help preserve existing knowledge. The ablations show that these components improve editing performance, but leave the preservation mechanism unresolved. An analysis linking subspace overlap and update norms to actual forgetting would strengthen the paper. Comparisons at matched editing efficacy or update norm would also help explain where the preservation benefits come from.

The method's sensitivity to edit selection and batch composition remains unclear. Defining Pq=QqQq⊤P_q=Q_qQ_q^\top, the update satisfies

ΔK1=RPq,R−ΔK1=R(I−Pq).\Delta K_1=RP_q,\qquad R-\Delta K_1=R(I-P_q).

Each edit therefore depends on a subspace determined by the entire batch, so the same edit may receive different updates depending on the other edits included. The cumulative S4S^4 criterion selects directions based only on the singular values of the keys, without considering the target residuals. Directions with small singular values may still carry information needed to distinguish different editing targets associated with similar keys. Discarding them could systematically hurt certain groups of edits even when average performance remains strong. I would like to see results across independently sampled edit sets, tests that keep an edit fixed while changing the other edits in its batch, and failure rates broken down by key similarity and relation type.

The trade-off between fitting more edits and preserving existing knowledge needs a clearer analysis as batch size grows. The retained rank satisfies

q≤rank⁡(CK1)≤min⁡(d−1,N),q\leq\operatorname{rank}(CK_1)\leq\min(d-1,N),

and the unfitted target residual is R(I−Pq)R(I-P_q). As editing requirements become more diverse, truncation may discard directions needed to fit some of the additional edits. Retaining more directions, however, may introduce smaller singular values or increase overlap with existing knowledge.

Appendix D reports 1,089 retained directions for 2,000 edits and 3,451 for 10,000 edits at layer 13 of GPT-2 XL. It would be useful to connect these measurements to editing failures and forgetting. The 10,000-edit results also cover only GPT-2 XL, with limited corresponding analysis of update norms and downstream performance. A useful experiment would keep a core set of edits and probes for existing knowledge fixed, then progressively add other edits to reach batch sizes of 2,000, 5,000, and 10,000. Tracking performance on the same examples, together with projection residuals, retained rank, and update norms across multiple models, would make the effects of increasing batch size much clearer.

The related-work discussion would benefit from including recent studies on large-scale editing and editing stability, particularly [1] and [2]. Discussing how the proposed approach relates to these studies would help clarify its contribution and the differences in how editing instability is addressed.

[1] EAMET: Robust Massive Model Editing via Embedding Alignment Optimization. ICLR 2026.

[2] Spectral Characterization and Mitigation of Sequential Knowledge Editing Collapse. arXiv 2026.

Questions for authors
  1. How do centering and truncated SVD help preserve existing knowledge beyond improving numerical stability?
  2. How sensitive is the method to edit selection and batch composition when target residuals rely on discarded directions?
  3. How does increasing batch size affect editing efficacy and knowledge preservation across models when the core edits and knowledge probes are held fixed?
4 comments

Thank you for your thoughtful feedback and for your positive assessment of our paper’s motivation and structure. Below, we respond to each of your concerns in turn.

[Weakness 1 & Question 1. Preservation]

We agree that constraining update directions alone does not establish that those directions are suitable for preserving existing knowledge. As your analysis shows, the effect on existing representations also depends on their overlap with the retained editing subspace. Our method does not explicitly control this overlap.

Centering and TSVD impose structural constraints on the update. Centering prevents changes along the all-ones input direction, yielding Δ1=0\Delta\mathbf{1}=0, while TSVD excludes discarded singular directions from the update and limits amplification associated with small singular values. These constraints differ from a preservation penalty constructed from external reference representations. Our method uses only the new editing keys for regularization.

To further examine the relationship between update norm and efficacy, we compared the centered TSVD update with a centered, SVD-based ridge update using GPT-2 XL on ZSRE for 5,000 edits. We measured fitting error and update norm at the first edited layer.

MethodEfficacyGeneralizationSpecificityFitting ErrorUpdate NormRuntime (min)
Ours70.7 ± 1.361.8 ± 1.023.0 ± 0.2610.1 ± 0.692.3 ± 0.128.4 ± 0.5
Ridge63.0 ± 2.354.1 ± 1.620.6 ± 0.4447.1 ± 1.2110.5 ± 0.229.2 ± 0.4

Our method achieves higher efficacy, generalization, and specificity with a smaller update norm. These results show that a larger update norm does not necessarily lead to higher efficacy in this setting.

We also evaluated a sequential extension, denoted Ours**. This extension retains centering and TSVD, adds ridge regularization to penalize the update magnitude in all directions, and incorporates accumulated statistics from previous edits to mitigate cumulative interference. These statistics are collected from previous edits rather than precomputed from an external reference corpus. We conducted sequential editing on ZSRE using GPT-2 XL with a batch size of 100 for N=1,000N=1,000.

MethodEfficacyGeneralizationSpecificity
MEMIT (w/ K0K_0)77.772.025.9
AlphaEdit (w/ K0K_0)92.986.825.6
Ours** (w/o K0K_0)87.681.525.3

These results indicate that the proposed method can be extended to sequential editing without relying on K0K_0. We have revised Figure 1 to more clearly illustrate how centering and truncated SVD help preserve existing knowledge.

[Weakness 2 & Question 2. Sensitivity]

We agree that the cumulative S4S^4 criterion depends only on the singular values of the keys and does not account for the target residuals. To examine sensitivity to edit selection and batch composition, we conducted experiments on MCF using GPT-2 XL. We held one edit (ID 0) fixed and randomly sampled the remaining 999 edits to construct a batch of 1,000 edits. We repeated this procedure three times with different batch compositions.

ScoreEfficacyGeneralizationSpecificityFluencyConsistency
85.398.992.370.3623.241.5
86.399.093.871.4619.940.6
86.298.494.071.4624.942.1

The fixed edit succeeded in all three runs. Following your suggestion, we also computed cosine similarities from the original key matrix K1K_1 at the first edited layer. For each edit, we calculated the maximum cosine similarity between its key and the other editing keys in the same batch. We pooled all evaluations across the three runs, sorted them by maximum similarity, and divided them into three equally sized groups.

SimilarityNSuccessFailureFailure Rate (%)
Low1,00099370.7
Medium1,00099820.2
High1,000972282.8

The high-similarity group had the highest failure rate among the three groups.

We also grouped the evaluations by relation type. The table below includes only relation types with at least one failure.

Relation typeNSuccessFailureFailure Rate (%)
P127686622.9
P131898544.5
P1384538715.6
P140757411.3
P17613012643.1
P2013413221.5
P2711010643.6
P276767422.6
P3012111921.7
P407333126.1
P41315715610.6
P4632018210.0
P641403825.0
P74011511321.7

Among these relation types, P138 had the highest failure rate, with 7 failures in 45 evaluations (15.6%).

[Weakness 3 & Question 3. Trade-off]

Following your suggestion, we held a core set of 100 edits and their associated existing-knowledge probes fixed while progressively adding other edits to the batch. We measured efficacy, generalization, specificity, and projection residuals on the same core set. Retained rank and update norm were measured at the first edited layer of each model. We evaluated batch sizes of 2,000, 5,000, and 10,000 for GPT-2 XL and GPT-J, and batch sizes of 2,000 and 5,000 for Llama 3.

ModelBatch SizeEfficacyGeneralizationSpecificityProjection ResidualRetained RankUpdate Norm
GPT-2 XL200082.274.424.676.7108853.4
GPT-2 XL500072.867.023.289.8234982.6
GPT-2 XL1000064.653.624.3109.9345091.1
GPT-J200096.794.627.3189.2107618.5
GPT-J500096.093.427.1319.5258129.3
GPT-J1000093.087.926.1505.7465839.7
Llama 3200085.280.131.118.410899.9
Llama 3500081.177.729.930.4263316.4

For each model, efficacy and generalization decreased as batch size increased. For GPT-2 XL, specificity did not decrease monotonically with batch size.

[Weakness 4. Related work]

EAMET [1] and REVIVE [2] address performance degradation in large-scale knowledge editing. EAMET [1] addresses misalignment between key and residual embeddings, whereas our method regularizes the closed-form update without relying on external preservation statistics. REVIVE [2] explicitly preserves dominant singular subspaces of original weights. Our method do not explicitly identify weight subspaces associated with existing knowledge.

Thank you again for your constructive feedback. We have included the additional experiments in Appendix D.1 and related work in Section 5 in the revised manuscript.

[1] Dai, Y., Ji, Z., Li, Z. and Wang, S., 2026. EAMET: Robust massive model editing via embedding alignment optimization. In International Conference on Learning Representations. [2] Zhang, C., Zhang, M., Ye, X., Cheng, R., Zhou, Z., Zhou, Y., Ren, P. and Chen, Z., 2026. Spectral characterization and mitigation of sequential knowledge editing collapse. In Annual Meeting of the Association for Computational Linguistics.

Thanks for the detailed responses, which address most of my concerns. Please make sure these additional results are included in the revised manuscript. I will maintain my original score.

We sincerely appreciate your positive feedback and follow-up. You can see the additional results in Appendix D.1 of the revised manuscript. Thank you for your detailed review, which has strengthened the paper.

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

The paper proposes a closed-form weight-update procedure for factual association editing that avoids constructing or storing external, corpus-derived representations of the pretrained knowledge to be preserved, denoted by K0K_0. Starting from a least-squares objective involving only the requested edits, the method computes an update of the form Δ=(V1−WK1)(PK1)k†\Delta=(V_1-WK_1)(PK_1)^{\dagger}_{k}, where P=I−11⊤/dP=I-\mathbf{1}\mathbf{1}^{\top}/d removes the within-vector mean and the subscript kk denotes a truncated pseudoinverse. The retained rank is selected using either spectral entropy or a cumulative fourth-power singular-value criterion, with the latter used by default. The resulting update modifies existing feed-forward projection weights without introducing an auxiliary editor or inference-time routing module. Experiments on ZSRE and Multi-CounterFact across GPT-2 XL, GPT-J, and Llama 3 demonstrate competitive batch-editing performance, including a 10,000-edit evaluation on GPT-2 XL, while additional Mistral, MQuAKE, and downstream evaluations reveal limitations under sequential updates and multi-hop reasoning. The efficiency study reports substantial savings under a protocol that amortizes baseline preprocessing over ten editing runs.

Critical strengths

Eliminating corpus-derived covariance precomputation and persistent storage while directly updating existing model weights removes a concrete deployment dependency without introducing adapters, routers, or separately trained auxiliary models.

Direct pseudoinversion avoids explicitly forming the Gram matrix that squares the nonzero-spectrum condition number, while truncated SVD suppresses small-singular-value amplification, and the component ablation separately examines pseudoinversion, truncation, and centering.

The evaluation combines multiple model architectures and editing scales with FastMEMIT and UltraEdit comparisons, multi-hop reasoning, and downstream capability probes, providing a broader assessment than target-fact recall alone.

Critical weaknesses

Although the appendix includes a direct-pseudoinverse baseline and truncation ablations, the distinctive benefit of the spectrum-only S4S^4 rank-selection rule is not isolated from ordinary regularization because no matched, K0K_0-free ridge-regularized solve is evaluated.

  • Experiment: On GPT-2 XL, compare the proposed centered TSVD update with a centered, SVD-based ridge update on batches of 100 and 1,000 ZSRE edits, keeping the edit requests, target-value optimization, and numerical precision identical. Select the ridge coefficient on a separate 100-edit calibration batch, and report efficacy, generalization, specificity, residual fitting error, update norm, and runtime over three seeds; this requires changing only the small linear-algebra component of the existing editing pipeline.

The proposed sequential variant reaches only 24.5% efficacy and 23.5% specificity after 1,000 Mistral edits, and the existing comparison does not disentangle accumulated interference from the additional first-principal-component removal used only for sequential editing.

The reported up-to-30× speedup includes 5–20 hours of baseline preprocessing amortized over just ten 100-edit runs, so it does not establish a similarly large advantage when covariance statistics are already cached or when using the reduced-precomputation FastMEMIT baseline.

  • On one backbone and the same GPU, could you please benchmark the proposed method, MEMIT, AlphaEdit, and FastMEMIT at batch sizes of 100 and 1,000, separating target-value optimization, covariance construction/loading, and the update solve. Measure warm-cache editing time directly and use the separately measured one-time preprocessing cost to report total amortized time for 1, 10, and 100 batches, alongside editing quality; this extends the existing timing study without requiring new model training.
Questions for authors

Please see above.

2 comments

Thank you for your thoughtful feedback and for recognizing the breadth of our evaluation. Below, we respond to each of your concerns in turn.

[Weakness 1. Ablation study]

Following your suggestion, we compared the proposed centered TSVD update with a centered, SVD-based ridge update using GPT-2 XL and ZSRE on an NVIDIA RTX 3090. We evaluated four batch sizes (100, 1,000, 2,000, and 5,000), keeping the edit requests, target-value optimization, and numerical precision identical. We selected a ridge coefficient of λ=10\lambda=10 based on the score and efficacy measured on a separate 100-edit calibration batch.

CoefficientScoreEfficacyGeneralizationSpecificity
0.549.596.087.325.8
149.896.288.126.0
249.296.290.025.3
549.095.089.525.3
1049.995.889.226.0
2049.695.487.426.0
5048.292.784.125.3
10049.989.584.027.0
20049.386.181.427.0
50045.474.964.826.8
100041.859.055.727.1

We report efficacy, generalization, specificity, residual fitting error, update norm, and runtime across three seeds. The fitting error is defined as ∥ΔK1−R∥F\|\Delta K_1-R\|_F, where K1K_1 denotes the uncentered key matrix. The fitting error and update norm are measured at the first edited layer.

Batch SizeMethodEfficacyGeneralizationSpecificityFitting ErrorUpdate NormRuntime (min)
100Ours97.8 ± 0.984.3 ± 2.724.3 ± 1.813.7 ± 0.914.5 ± 0.30.37 ± 0.0
100Ridge98.4 ± 0.484.7 ± 2.524.3 ± 1.715.1 ± 1.013.5 ± 0.20.40 ± 0.0
1,000Ours86.3 ± 0.375.1 ± 0.525.3 ± 0.4131.7 ± 5.848.7 ± 0.84.7 ± 0.1
1,000Ridge85.4 ± 0.673.2 ± 0.324.9 ± 0.4100.9 ± 3.646.4 ± 0.34.3 ± 0.5
2,000Ours81.1 ± 0.770.4 ± 0.724.9 ± 1.2250.8 ± 2.270.3 ± 0.49.2 ± 0.4
2,000Ridge77.6 ± 2.366.0 ± 1.424.2 ± 1.3187.0 ± 2.268.7 ± 0.29.3 ± 0.4
5,000Ours70.7 ± 1.361.8 ± 1.023.0 ± 0.2610.1 ± 0.692.3 ± 0.128.4 ± 0.5
5,000Ridge63.0 ± 2.354.1 ± 1.620.6 ± 0.4447.1 ± 1.2110.5 ± 0.229.2 ± 0.4

The proposed centered TSVD update achieves higher editing performance than the centered, SVD-based ridge update for N=1000N=1000. These results support the effectiveness of the proposed update at larger batch sizes.

[Weakness 2. Sequential editing]

We introduced first-principal-component removal to improve performance in the sequential-editing setting.

MethodEfficacyGeneralizationSpecificity
UltraEdit68.862.547.8
Ours*24.523.123.5
Ours0.40.21.3

Here, Ours* removes the first principal component, whereas Ours does not. The results show that removing this component improves all three metrics. We have clarified this limitation in Appendix A and included this result in Table 8 in the revised manuscript.

[Weakness 3. Efficiency analysis]

We compared the proposed method, MEMIT, AlphaEdit, and FastMEMIT using GPT-2 XL on ZSRE with batch sizes of 100 and 1,000. All experiments were conducted on an NVIDIA RTX A6000. We report warm-cache editing time, target-value optimization time, covariance construction/loading time, and update solve time. We also report total amortized runtime for 1, 10, and 100 batches, alongside editing quality. The covariance construction time is excluded from the reported warm-cache editing time.

Batch SizeMethodEditing Time (min)Optimization (min)Covariance (min)Update Solve (s)
100MEMIT2.6 ± 0.02.2312.42.1
100AlphaEdit2.8 ± 0.02.2312.80.5
100FastMEMIT2.6 ± 0.02.22.52.1
100Ours2.5 ± 0.02.20.00.3
1,000MEMIT25.7 ± 0.122.5312.43.7
1,000AlphaEdit28.5 ± 0.022.5312.80.5
1,000FastMEMIT25.7 ± 0.022.53.03.7
1,000Ours25.8 ± 0.022.50.07.1
Batch SizeMethodTotal Time (1, min)Total Time (10, min)Total Time (100, min)EfficacyGeneralizationSpecificity
100MEMIT314.9338.1569.580.477.125.4
100AlphaEdit315.6340.9594.093.587.125.7
100FastMEMIT5.028.3260.684.579.425.4
100Ours2.525.3253.391.986.827.0
1,000MEMIT338.0569.22,880.680.874.026.0
1,000AlphaEdit341.3597.53,159.994.387.325.6
1,000FastMEMIT28.7260.12,573.583.476.226.0
1,000Ours25.8258.02,580.089.381.525.3

Our method has a warm-cache runtime comparable to those of MEMIT and FastMEMIT. It avoids the one-time precomputation cost. However, the advantage becomes less pronounced as the cost is amortized over more batches. We also report warm-cache editing times for batch sizes of 1, 10, and 100 edits.

Batch SizeMethodRuntime (s)EfficacyGeneralizationSpecificity
1MEMIT4.8 ± 0.075.075.066.7
1AlphaEdit3.1 ± 0.0100.0100.066.7
1FastMEMIT5.4 ± 0.0100.075.066.7
1Ours1.9 ± 0.0100.0100.066.7
10MEMIT19.6 ± 0.081.575.531.8
10AlphaEdit18.9 ± 0.1100.088.031.8
10FastMEMIT20.1 ± 0.086.079.531.8
10Ours16.6 ± 0.0100.0100.031.8
100MEMIT154.3 ± 0.280.477.125.4
100AlphaEdit168.8 ± 0.493.587.125.7
100FastMEMIT154.9 ± 0.584.579.425.4
100Ours152.0 ± 1.191.986.827.0

Thank you again for your constructive feedback. We have included these experiments in Appendices B.1 and C.1 in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.