acceptodds
Under review as a conference paper at ICLR 2027

CanonEdit: Canonical Quotient Editing

Abstract

Sequential factual editing asks models to absorb requested facts over a long update stream while preserving behavior on the remaining knowledge. Yet each target specifies what behavior should change, not which parameter update should realize it. Many updates can write the same fact, but their side effects differ and can accumulate; the update retained at each step therefore shapes how long the model remains editable. We introduce **CanonEdit**, which groups updates with the same target action into quotient-space equivalence classes, measures their cost on history and unrelated prompts, and commits the least-cost representative that passes nonlinear success and protection-budget checks. This quotient-space representative selection is a target-constrained Bregman projection with a closed-form KKT solution. We prove that CanonEdit selects the successful local update with the smallest predicted protected change; under a stated budget, its feasible set retains every feasible hard-null update in the same trust region and admits additional bounded-cost directions. Across three datasets and four backbone families, seven 10,000-request streams achieve 99.24-100% online writing efficacy, 0-76 failed writes, and 99.97-100% measured specificity. These results show that quotient-space representative selection can control cumulative interference and keep models receptive to new knowledge over long update streams.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Model editing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_wxQYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the sequential editing task. The authors seek to find, among the infinitely many feasible solutions that achieve the target edit, a parameter modification that is as good as possible. To this end, they propose Canon Edit. The method first employs a linear approximation in the parameter space to delineate the feasible region satisfying the editing objective, which serves as a constraint. It then constructs a second-order approximation of a cost function to measure the degree of disruption to previously stored knowledge. A correction mechanism is further introduced to refine the solution when the second-order approximation is inaccurate. This formulation transforms the final optimization task into a strictly convex optimization problem that admits a closed-form optimal update. Intuitively, the overall framework resembles a trust-region optimization process in which editing success is treated as a constraint while locality (preservation of unrelated knowledge) is treated as the optimization objective. The authors demonstrate through experiments that the proposed method indeed achieves strong performance.

Critical strengths
  • The idea introduced in this paper is very interesting. The constrained optimization perspective of knowledge editing in this paper provides an elegant framework that is valuable for future work to reference and build upon.
  • The experimental evaluation is thorough, extending the testing range to 10,000 sequential edits. At this scale, the proposed method indeed demonstrates strong and consistent performance.
Critical weaknesses
  • Canon Edit directly formulates the minimization of disruption to unrelated knowledge as the sole optimization objective and solves for a unique solution under the editing constraint. However, real-world editing tasks involve additional important desiderata beyond faithfulness and locality; for example, improving the portability on the edited knowledge. If such objectives conflict with the current constraint, it is unclear how they can be reconciled. These additional objectives do not appear to lend themselves to simple second-order approximations and are thus difficult to integrate into the proposed framework. Moreover, the current closed-form solution architecture seemingly provides no mechanism to adjust the trade-off among different objectives.
  • The core of Canon Edit is built upon numerous approximation steps. Given the highly complex and nonlinear parameter space of LLMs, the accumulated approximation error is inevitably substantial. This makes the method heavily dependent on the correction mechanism. The effectiveness and robustness of the overall approach may therefore be sensitive to how well the correction procedure compensates for these approximation errors.
Questions for authors
  • The main experimental results in Table 1 appear somewhat unusual. For instance, why are the online efficacy and generalization metrics (Rel. & Gen.) for AlphaEdit so low (e.g., 25.17% Rel.), even performing below one-shot methods like ROME and MEMIT? Meanwhile, the Locality scores for these methods are all very high, which collectively gives the impression that the edits were not successfully applied in the first place. I would appreciate it if the authors could further clarify whether the baseline implementations and hyperparameter selection processes differ from their standard/official implementations, or explain other possible reasons that may account for this.

I would be happy to increase my score if the authors clarify my concerns.

4 comments
AuthorsOpenReview ↗

Response to Reviewer wxQY

We thank the reviewer for recognizing the value of CanonEdit's constrained optimization perspective and its consistent performance over 10,000 sequential edits. Your comments raise important questions about objective flexibility, approximation reliability, and baseline evaluation. We address these concerns below.

1. Extending CanonEdit to Multiple Editing Objectives and Controllable Trade-offs (W1)

Thank you for highlighting portability and the need to balance multiple editing objectives. CanonEdit provides a natural mechanism for incorporating these requirements by separating behavioral constraints from the protected-cost objective.

Additional differentiable requirements, such as portability across related queries, can be expressed as target-margin constraints:

min⁡z12z⊤Bzs.t.Aeditz≥bedit,Aportz≥bport.\min_z \frac{1}{2}z^\top Bz \quad \text{s.t.}\quad A_{\mathrm{edit}}z\geq b_{\mathrm{edit}},\quad A_{\mathrm{port}}z\geq b_{\mathrm{port}}.

When the local feasible set is nonempty, the resulting convex quadratic program retains the same optimization structure. Nonlinear verification subsequently evaluates the realized editing and portability behavior.

For multiple protection objectives, CanonEdit also admits a weighted local metric:

B=∑iλiGi+αM,λi≥0,B=\sum_i\lambda_iG_i+\alpha M, \qquad \lambda_i\geq0,

where each GiG_i represents a locally quadratic protection criterion. The weights λi\lambda_i control their relative importance, while the target margins γ\gamma and protected-disturbance budget ϵ\epsilon regulate feasibility and acceptance.

These extensions clarify how CanonEdit accommodates additional behavioral requirements and explicit objective trade-offs. They preserve the local convex formulation when the corresponding constraints and cost approximations satisfy its assumptions. The reported experiments establish the original faithfulness–locality setting, providing the foundation for these broader formulations.


2. Reliability of Local Approximations and Nonlinear Refinement (W2)

We appreciate your concern regarding approximation errors in nonlinear LLMs. CanonEdit addresses this through a combination of local geometric estimation, full-model verification, and targeted re-linearization. We provide additional empirical evidence for each component.

First, the original GPT-J mechanism analysis (Fig. 3(c)) reports a 0.74 log-scale correlation (95% bootstrap CI: [0.65, 0.82]) between predicted protected energy and measured disturbance. Our new shared-pool experiment further evaluates generalization of the protected-cost ranking: CanonEdit achieves a held-out static MAE of 0.2504, compared with 0.4105, 0.3493, and 0.3482 for max-margin, first-valid, and random-valid selection.

Second, nonlinear verification directly checks candidate feasibility in the original model. In the 10K CounterFact–GPT-J experiment, 93.90% of requests follow the closed-form path, 6.13% trigger re-linearization, and the average number of full candidate evaluations is 1.40. These measurements characterize how frequently the method requires nonlinear correction.

To further examine robustness, we conducted a controlled GPT-2-XL CounterFact experiment (N=600), varying the maximum number of refinement steps KK:

KOnline Target (%) ↑Hist. (%) ↑Protected Drift ↓Failed Writes ↓Time (s/edit) ↓
089.8368.170.1055651.445
192.8369.500.1097501.514
290.5070.330.0983681.388
493.6770.170.1141441.802

Increasing the refinement budget from K=0K=0 to K=4K=4 improves online target success by 3.84 percentage points and reduces failed writes from 65 to 44. Meanwhile, K=2K=2 achieves the lowest protected drift, illustrating the interaction between correction strength, interference, and computational cost.

Together, these results establish that the local protected-cost approximation provides informative candidate rankings, while nonlinear refinement offers a measurable and controllable mechanism for recovering target feasibility. The limited re-linearization frequency in the original 10K experiment further supports the practical efficiency of this design.


3. Baseline Fairness and the Unexpected AlphaEdit Results (Q1)

Thank you for carefully examining the AlphaEdit results in Table 1. Table 1 reports 25.17% online efficacy for AlphaEdit on CounterFact-plus/OneHop with GPT-J, motivating a closer examination of target construction and evaluation settings.

Table 1 evaluates the CounterFact-plus/OneHop request stream under the reported controlled sequential-editing protocol, including a fixed GPT-J editing layer and a teacher-forced target-versus-original likelihood criterion. AlphaEdit obtains 25.17% online efficacy, while AlphaEdit with the Forward Replay target achieves 88.32% on that same stream.

The substantial efficacy increase with the Forward Replay target highlights the importance of target construction under the controlled sequential protocol. The comparison specifies the editing layer, evaluation criterion, and request stream to support reproducibility.

We traced the higher AlphaEdit+FR efficacy in Table 15 (93.56%) to the distinct CounterFact/MCF request stream. Both tables now identify their exact streams, accounting for the different reported values. The 25.17% AlphaEdit result is measured under the specified OneHop protocol.

Our shared-pool experiment additionally compares selection rules using identical parent weights, verified candidates, and evaluation conditions. CanonEdit achieves the lowest held-out disturbance (0.2504) across all four selectors, with strictly positive paired 95% confidence intervals against each alternative.

These matched-condition results establish the independent contribution of canonical representative selection and complement the sequential baseline comparisons.

We sincerely appreciate the reviewer's constructive feedback and encouragement. The theoretical clarifications and additional experiments further demonstrate how CanonEdit combines constrained optimization, informative local protection geometry, and efficient nonlinear verification to identify low-interference successful updates. We hope these additional results satisfactorily address your concerns and provide sufficient grounds for reconsidering your initial rating.

AuthorsOpenReview ↗

Follow-up on Rebuttal and Further Discussion

Dear Reviewer wxQY,

Thank you again for your encouraging feedback and helpful suggestions. In our rebuttal, we clarified how CanonEdit can support multiple editing objectives, added experiments on nonlinear refinement, and explained the AlphaEdit results and evaluation settings.

As the ICLR discussion period is approaching its end, we would sincerely appreciate your thoughts on whether these responses address your concerns. We would be happy to discuss any remaining questions.

Thank you very much for your time and consideration.

AuthorsOpenReview ↗

CanonEdit (#5375): Request for Response from Reviewer wxQY

Dear Reviewer wxQY,

Thank you again for your encouraging comments and for expressing your willingness to reconsider your assessment.

As the discussion period is nearing its end, we sincerely hope to hear your thoughts on our rebuttal, particularly our clarifications of the AlphaEdit baseline results and the additional experiments on nonlinear refinement.

We would greatly appreciate it if you could review these responses and let us know whether your main concerns have been addressed.

Thank you very much for your time and consideration.

Reviewer_BZB1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This article introduces CanonEdit, exploring how to select parameter updates that cause minimal interference with existing knowledge from multiple successful updates in continuous knowledge editing. The method characterises updates that have the same effect through quotient space equivalence classes, uses Bregman projection with target constraints to pick low-cost representatives, and combines nonlinear verification to ensure the edit actually succeeds.

Critical strengths

1.The paper formulates update selection in sequential editing as constrained optimization, balancing editing success and knowledge preservation.

2.Extensive experiments across multiple models and datasets, including 10,000 sequential edits, demonstrate the method’s effectiveness.

Critical weaknesses

1.The local formulation is a standard constrained quadratic program. In the single-basis Riesz setting, the cost matrix does not affect the optimum under fixed target constraints, limiting evidence for canonical selection from this setting.

2.Final-10K retention remains unclear despite high online efficacy. CounterFact–GPT-J AlphaEdit+FR results differ between Tables 1 and 15; Appendix F describes evaluating edits 4,501–5,500 at a 5K checkpoint. Please clarify these discrepancies and extend the retention audit to 10K.

Questions for authors

1.Could the authors supplement pairwise likelihood comparisons with token-level top-1 accuracy used in ROME/MEMIT-style zsRE evaluation, and report CounterFact Fluency and Consistency alongside baselines under the same standard protocol?

2.The description appears to reuse paraphrases for optimization and evaluation. Could the authors clarify the split and report held-out paraphrase results, distinct from the held-out protection audit?

3.Beyond the reported ablations and timing, could the authors isolate canonical selection by holding candidate construction, search, and verification budgets fixed?

4 comments
AuthorsOpenReview ↗

Response to Reviewer BZB1

We thank the reviewer for recognizing the value of our constrained optimization perspective and extensive sequential-editing experiments. Your comments have helped us sharpen the theoretical interpretation of canonical selection and strengthen the evaluation of its independent contribution, generalization, and historical retention. We address the three main concerns below.

1. Canonical Selection: Theoretical Novelty and Controlled Validation (W1 & Q3)

Thank you for highlighting the distinction between standard quadratic programming and the underlying editing contribution. Our central insight is that multiple parameter updates can satisfy the same editing requirements while inducing different interference with protected knowledge. CanonEdit formalizes this ambiguity through target-equivalence classes and selects low-interference representatives using target-constrained Bregman projection and nonlinear verification.

Your observation about the single-basis Riesz setting is mathematically correct. For a one-dimensional chart with one active constraint az≥b>0az\geq b>0, assuming a>0a>0, the optimum is

z∗=ba,z^*=\frac{b}{a},

independent of the positive scalar cost BB. This clarifies the role of protected geometry: multi-coordinate projection determines the minimum-cost update within a chart, while canonical selection compares protected costs across distinct verified candidates. Theorem 1 characterizes the local optimum, and the verified-atlas procedure implements the broader selection principle.

To directly validate this contribution, we conducted a controlled shared-pool experiment on a 600-request GPT-2-XL CounterFact stream. Among 456 requests with feasible candidate pools, 335 contain multiple verified candidates. All selectors operate on identical parent weights, candidate pools, search results, and nonlinear verification conditions.

SelectorHeld-out MAE ↓In-pool Drift ↓
CanonEdit0.25040.2435
Max-margin0.41050.3759
First-valid0.34930.3228
Random-valid0.34820.3321

On the 335 multi-candidate requests, CanonEdit reduces held-out disturbance by 0.2179 relative to max-margin (95% CI: [0.1535, 0.2959]), 0.1346 relative to first-valid ([0.0810, 0.2001]), and 0.1331 relative to random-valid ([0.0831, 0.1953]).

All compared candidates satisfy the same nonlinear target and protection gates. The matched experimental conditions isolate the effect of representative selection. These findings directly establish the contribution beyond the local QP solution: canonical selection consistently identifies lower-interference updates among successful candidates, with measurable benefits on independently evaluated protected behavior.


2. Evaluation Reliability: Independent Generalization and Prediction Accuracy (Q1 & Q2)

We appreciate your suggestions regarding independent paraphrase evaluation and standard prediction-level metrics. We conducted a strict construction/evaluation split and extended the evaluation to teacher-forced token-level Top-1 accuracy and greedy-generation exact match.

The auxiliary manifest covers 10,000 request records, with zero exact prompt overlap between construction and evaluation paraphrase, history, and static fields. Evaluation-only prompts are excluded from candidate construction, nonlinear verification, and selection.

The strict GPT-2-XL CounterFact experiment (N=600) yields:

Evaluation Metric (%)CanonEditNo-canonical
Pairwise target efficacy97.3397.33
Held-out paraphrase success57.8361.17
Target token Top-19.992.22
Held-out paraphrase Top-19.633.39
Greedy target EM2.000.67

At identical pairwise target efficacy, CanonEdit improves target Top-1 by 7.77 percentage points, held-out paraphrase Top-1 by 6.24 points, and greedy exact match by 1.33 points. These results demonstrate that canonical selection improves strict prediction accuracy and provide a more complete characterization of editing outcomes.

We additionally evaluate protected behavior on independent static prompts:

Held-out Protection MetricCanonEditNo-canonical
Protection retention (%)55.4756.25
Protection MAE ↓6.1764.723

These measurements distinguish paraphrase generalization from unrelated-prompt protection and reveal their respective trade-offs. The shared-pool experiment in Section 1 further establishes CanonEdit's protection advantage under matched candidate pools and verification conditions.

Together, the independent evaluation and strict prediction metrics strengthen the empirical assessment of editing efficacy and generalization. These CounterFact measurements provide complementary evidence on held-out generalization, token-level accuracy, and generation, with each metric reported under its specified protocol.

3. Long-Horizon Retention and Experimental Consistency (W2)

Thank you for emphasizing the importance of final-checkpoint retention and consistency across reported results. We conducted an additional exhaustive historical-retention audit, evaluating every preceding edit at checkpoints 200, 400, and 600 in the strict GPT-2-XL CounterFact experiment.

All-Prior Retention (%)@200@400@600
CanonEdit target74.0063.7566.33
No-canonical target66.0068.7568.83
CanonEdit paraphrase72.0060.5062.33
No-canonical paraphrase55.5060.5057.67

Both methods achieve 97.33% online target efficacy. At checkpoint 600, CanonEdit improves prior-paraphrase retention by 4.66 percentage points, while no-canonical achieves 2.50 points higher prior-target retention.

These results provide a direct measurement of accumulated factual retention and clarify the relationship between current editing efficacy, prior-target preservation, and paraphrase retention. The original 10K trajectories characterize online editing performance, and the new 600-step audit quantifies retention across all prior edits at the specified checkpoints.

Thank you for drawing attention to the Table 1 and Table 15 results. We traced the AlphaEdit+FR entries to two distinct completed GPT-J-6B request streams:

Metric (%)Table 1: CounterFact-plus/OneHopTable 15: CounterFact/MCF
Rel./Eff.88.3293.56
Gen.67.3874.92
Hist./Ret.60.7362.27

The revised tables identify the two request streams as CounterFact-plus/OneHop (Table 1) and CounterFact/MCF (Table 15), preserving their recorded measurements.

We also corrected the Appendix F cohort description: at the common 5K checkpoint, the audit replays 200 uniformly sampled requests from edits 1–1,000 and 100 from edits 4,501–5,000. Table 10 now labels the cohorts Early (n=200) and Recent (n=100), preserving the original measurements.

These corrections make the provenance of both 10K experiments and the 5K replay cohorts explicit; the 600-step all-prior audit is reported with its own checkpoint protocol.

We sincerely appreciate the reviewer's thoughtful suggestions. The theoretical clarifications and additional controlled experiments strengthen the evidence for CanonEdit's central contribution: target-constrained representative selection provides a principled and empirically validated mechanism for reducing interference among successful updates. We hope these results help resolve your concerns regarding novelty and empirical validation and support a more favorable assessment of our work.

AuthorsOpenReview ↗

Follow-up on Rebuttal and Further Discussion

Dear Reviewer BZB1,

Thank you again for your constructive suggestions. We added a controlled comparison to isolate the benefit of canonical selection, evaluated editing performance on independent prompts, and clarified the historical retention results and differences between Tables 1 and 15.

As the ICLR discussion period is approaching its end, we would be grateful to hear your thoughts on these additions and welcome any further discussion.

Thank you very much for your time and consideration.

AuthorsOpenReview ↗

CanonEdit (#5375): Request for Response from Reviewer BZB1

Dear Reviewer BZB1,

As the discussion period is approaching its end, we sincerely hope to have the opportunity to discuss our rebuttal with you.

We have provided additional experiments and clarifications addressing your concerns about canonical selection, independent evaluation, and historical retention.

We would be very grateful if you could take a moment to review our responses and share your thoughts on whether they address your concerns.

Thank you very much for your time and consideration.

Reviewer_izyNOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents CanonEdit, a method for sequential knowledge editing that aims to reduce interference with existing knowledge and preserve editing performance as more edits are applied. The method groups parameter updates that achieve the same editing objective into equivalence classes, then uses a target-constrained Bregman projection to find an update that minimizes interference with protected knowledge. The resulting local optimization problem admits a closed-form solution, and a nonlinear validation step is used to check the feasibility of the edit. Experiments on three datasets and four models suggest that CanonEdit can maintain high editing success rates over sequences of up to 10,000 edits while reducing changes to protected knowledge.

Critical strengths

1. An Interesting Problem Perspective: The paper highlights an important aspect of sequential knowledge editing: parameter updates that achieve the same editing objective can have different effects on existing knowledge. This observation motivates the proposed constrained optimization framework.

Critical weaknesses

1.The core motivation lacks sufficient validation: Section 4 (Lines 140–142) takes as its core premise that "The success editing in Section 3 does not uniquely determine a parameter update: many updates can favor the new answer while affecting protected facts differently." However, no experiments are introduced to verify the reliability of this hypothesis, which is thus insufficient to adequately support the motivation of the proposed method.

2. Test data leakage: Eq. (5) incorporates paraphrase prompts (D_t^{para}) into the editing constraint, while Generality is evaluated based on paraphrases; similarly, Eq. (3) uses unrelated prompts (D_t^{stat}) to optimize protected behavior, while Specificity is also evaluated based on such prompts.

Questions for authors

1. Missing Strong Theoretically Related Baselines: The paper lacks comparisons with several strong and closely related methods, including SPHERE, which identifies low-interference update directions through hyperspherical geometry; REVIVE, which filters harmful updates in the spectral space of weight matrices; and CrispEdit, which similarly leverages Bregman divergence and local curvature information to preserve model capabilities. Comparisons with these methods are necessary to establish the proposed approach's effectiveness and novelty.

2.Overly Simplistic Evaluation Metrics.: CanonEdit evaluates editing success solely by whether the target answer has a higher likelihood than the original answer. Although this follows the conventional CounterFact metric, it does not guarantee correct predictions or generations. Prior work (e.g., MEMIT) has demonstrated that strong likelihood-based editing scores can coexist with severe generation failure. Moreover, ZsRE baselines commonly employ Top-1 prediction accuracy. The authors should additionally report teacher-forced Top-1 accuracy (target-answer token-level matching under greedy decoding) to validate the claimed editing efficacy.

3. Tradeoff of Rel. vs. Hist. and Long-horizon experiment: Is the Hist. metric proposed for the first time in this paper? Please provide the detailed calculation method of Hist., how it is constructed, why only two entries are needed, and an example. I am not entirely clear about what Hist. specifically evaluates. Is it equivalent to Prior-Edit Retention? In typical sequential editing processes, whether previously edited samples can be correctly retained is one of the core evaluation metrics. In the following, I will treat Hist. as Prior-Edit Retention for discussion (if Hist. does not have the following meaning, please clarify the necessity of this metric and still supplement the experiments on Prior-Edit Retention below): Taking GPT-J-6B on ZsRE in Table 1 as an example, although CANONEDIT has better online efficacy than AlphaEdit or AlphaEdit (FR target), its Hist. is lower. This intuitively suggests that CANONEDIT sacrifices the accuracy of previously edited knowledge to improve the current editing effect. Please present, across multiple models (not only GPT-J-6B) and multiple datasets, the Hist. obtained by CANONEDIT and baseline methods on all previously edited samples at every 1K/2K/5K/10K edits (Note that the evaluation should be conducted on all previously edited samples. For example, at the 5K checkpoint, all 5K previously edited samples must be evaluated, rather than sampling only two.), shown as line charts to display the trends, rather than only showing the advantageous Rel. as in Figure 2.

4. Unclear Symbol Definitions and Descriptions: (1) Unclear symbol definitions lead to confusing narrative: In Lines 146–150, what specific process does "the search" refer to? The meanings of "changes" and E are not clearly specified (although Section 5 later implies they denote parameter updates, readers are easily confused when reading the earlier text). These symbols do not appear in the cited ROME/MEMIT papers, and their vague definitions here create comprehension difficulties. Furthermore, in Eq. (3), ξ is sampled from D_{C,t}, but the sampling source of the monitored protected-answer log-likelihoods r_t in Eq. (4) is not explicitly stated; (2) The lack of logical textual transitions, compounded by an excessive and dense accumulation of formulas, makes the paper difficult to read; (3) Figure-text error: In Figure 1, "argmin" is misspelled as "agrmin"; (4) The abbreviations of the metrics are inconsistent between the main text and the appendix, e.g., Rel. and Eff., Hist. and Ret., etc.

6 comments
AuthorsOpenReview ↗

Response to Reviewer izyN

We thank the reviewer for recognizing the significance of our problem perspective and for the constructive suggestions. Motivated by your comments, we conducted additional controlled experiments to validate the central hypothesis, examine generalization on independent prompts, and assess prediction accuracy and historical retention. We address each concern below.

1. Empirical Validation of the Core Motivation (W1)

Thank you for emphasizing the empirical foundation of our motivation. Our central insight is that multiple parameter updates can satisfy the same editing requirements while inducing substantially different changes to protected knowledge. We directly investigate this phenomenon through a controlled shared-pool experiment.

On a 600-request GPT-2-XL CounterFact stream, 456 requests yield verified feasible candidate pools, including 335 with multiple successful candidates. Four selectors share identical parent weights, candidate pools, and nonlinear verification conditions. We additionally measure interference on an evaluation-only static probe.

SelectorHeld-out MAE ↓In-pool Drift ↓
CanonEdit0.25040.2435
Max-margin0.41050.3759
First-valid0.34930.3228
Random-valid0.34820.3321

Among the 335 multi-candidate requests, CanonEdit reduces held-out disturbance by 0.2179 relative to max-margin (95% CI: [0.1535, 0.2959]), 0.1346 relative to first-valid ([0.0810, 0.2001]), and 0.1331 relative to random-valid ([0.0831, 0.1953]).

All compared candidates satisfy the same nonlinear target and protection gates. The matched experimental conditions isolate representative selection as the source of the observed differences. These findings directly substantiate our motivation: successful updates exhibit materially different collateral effects, and canonical selection consistently identifies lower-interference representatives on independently evaluated protected behavior.


2. Independence of Optimization and Evaluation (W2)

Thank you for raising the question of evaluation independence. We strengthened the empirical assessment through a strict construction/evaluation partition, allowing generalization and protection to be measured on previously unseen prompts.

The auxiliary manifest covers 10,000 request records, with zero exact overlap between construction and evaluation paraphrase, history, and static fields. Evaluation-only prompts remain excluded from candidate construction, nonlinear verification, and selection.

The strict GPT-2-XL CounterFact experiment (N=600) yields:

Evaluation-only MetricCanonEdit
Paraphrase success (%)57.83
Protection retention (%)55.47
Protection MAE ↓6.176

The shared-pool experiment further examines protection transfer. Two static probes participate in candidate construction and verification, and a third independently evaluates the resulting updates. CanonEdit achieves a held-out MAE of 0.2504, compared with 0.4105 for max-margin, 0.3493 for first-valid, and 0.3482 for random-valid.

These results establish measurable generalization and protection on independent evaluation prompts. The revised evaluation presents construction-set and held-out results with explicit protocol and dataset identifiers.


3. Comparisons with Related Geometric Editing Methods (Q1)

We appreciate your suggestions regarding SPHERE, REVIVE, and CrispEdit. These studies provide valuable geometric, spectral, and curvature-aware perspectives on controlling interference during model editing.

CanonEdit advances this direction by introducing target-constrained representative selection as an explicit optimization principle. The quotient-space formulation characterizes alternative parameter realizations satisfying the same target requirements, while the Bregman projection and nonlinear verification identify low-interference successful updates.

The controlled experiment in W1 directly establishes the effectiveness of this selection principle. With identical candidate pools and verification conditions, CanonEdit reduces held-out disturbance by 0.2179, 0.1346, and 0.1331 relative to max-margin, first-valid, and random-valid selection, respectively. All three improvements have strictly positive 95% confidence intervals.

These findings provide direct empirical evidence that canonical representative selection independently improves interference control, establishing a concrete methodological contribution alongside existing geometric and curvature-aware editing approaches. This controlled evidence complements our technical comparison with the related geometric and curvature-aware editing methods.

4. Evaluation Beyond Pairwise Likelihood (Q2)

Thank you for emphasizing the importance of prediction-level evaluation. We extended the empirical analysis to teacher-forced token-level Top-1 accuracy and greedy-generation exact match, providing a more detailed assessment of editing outcomes.

Under the strict GPT-2-XL CounterFact protocol, we obtain:

Metric (%)CanonEditNo-canonical
Pairwise target efficacy97.3397.33
Target token Top-19.992.22
Held-out paraphrase Top-19.633.39
Greedy target EM2.000.67

At identical pairwise efficacy, CanonEdit improves target token Top-1 by 7.77 percentage points, held-out paraphrase Top-1 by 6.24 points, and greedy exact match by 1.33 points.

These results reveal an additional benefit of canonical selection: the selected updates improve strict prediction accuracy while maintaining the same pairwise editing success. Reporting all three metrics also provides a clearer characterization of target preference, token-level prediction, and full-answer generation. These CounterFact results extend the original evaluation with complementary prediction-level evidence.


5. Definition and Construction of Hist. (Q3a)

Thank you for requesting a clearer explanation of historical retention. We distinguish two complementary measurements that capture historical knowledge preservation at different levels.

Historical-context Retention (`Hist.`) evaluates designated historical-context probes during sequential editing. Each request includes up to two such probes, providing a fixed-cost online measurement of interference with previously established behavior.

All-Prior Retention evaluates every preceding edit at checkpoint kk:

Rprior(k)=1k∑i=1k1[mi(xi;θk)>0].R_{\mathrm{prior}}(k)=\frac{1}{k}\sum_{i=1}^{k}\mathbf{1}[m_i(x_i;\theta_k)>0].

Here, mim_i measures the target-answer margin of edit ii under the model after kk updates. The evaluation includes all preceding requests.

These complementary measurements characterize immediate historical-context preservation and accumulated factual retention, respectively. The two-probe online protocol supports efficient monitoring throughout long editing streams, while checkpoint-based all-prior evaluation provides a comprehensive historical audit.


6. Online Efficacy and Long-Horizon Prior-Edit Retention (Q3b)

We appreciate your emphasis on the relationship between online writing efficacy and preservation of previously edited knowledge. We conducted an additional full-prior audit to quantify this relationship directly.

At checkpoints 200, 400, and 600, we evaluate every preceding edit in the strict GPT-2-XL CounterFact experiment.

All-Prior Retention (%)@200@400@600
CanonEdit target74.0063.7566.33
No-canonical target66.0068.7568.83
CanonEdit paraphrase72.0060.5062.33
No-canonical paraphrase55.5060.5057.67

Both methods achieve 97.33% online target efficacy. At checkpoint 600, CanonEdit improves prior-paraphrase retention by 4.66 percentage points, with prior-target retention of 66.33% versus 68.83% for no-canonical.

The results reveal distinct retention patterns across rewrite targets and their paraphrases, providing a more detailed understanding of the efficacy–retention relationship. In particular, CanonEdit demonstrates stronger preservation of previously edited paraphrase behavior at matched online efficacy.

The original 10K experiments characterize sustained online writability across multiple datasets and backbones; the new 600-step audit additionally measures retention over all preceding edits at each reported checkpoint. Together with the shared-pool findings, these results provide evidence linking canonical selection, immediate interference control, and historical knowledge preservation.

7. Notation and Presentation Clarifications (Q4)

Thank you for the detailed presentation suggestions. We have organized the methodological explanation around the progression from target-responsive update construction to verified representative selection.

The search comprises candidate-chart construction, within-chart coefficient optimization, and selection among verified candidates. Each Et,ℓE_{t,\ell} denotes a structured low-rank parameter-update basis, and zℓz_\ell determines its contribution:

Δθt(z)=∑ℓ=1dtzℓEt,ℓ.\Delta\theta_t(z)=\sum_{\ell=1}^{d_t}z_\ell E_{t,\ell}.

The quantity rtr_t denotes the number of monitored protected-answer scores associated with DC,t=Dthist∪DtstatD_{C,t}=D_t^{\mathrm{hist}}\cup D_t^{\mathrm{stat}}.

The complete method follows five connected stages:

Target-responsive update directions → Linearized target constraints → Protected-cost Bregman projection → Nonlinear verification → Minimum-cost representative selection.

The presentation consistently uses `Rel./Eff.` for online pairwise efficacy, `Hist./Ret.` for historical-context retention, and `All-Prior Retention` for checkpoint-based evaluation of all previous edits. Figure 1 uses the corrected `argmin` notation.

We sincerely thank the reviewer for the constructive feedback. The additional controlled experiments provide direct evidence for CanonEdit's central motivation: among multiple successful parameter updates, canonical representative selection significantly reduces interference on independently evaluated protected behavior. We hope these new findings and clarifications address your principal concerns and encourage you to reconsider your assessment of our work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary: Key Revisions and Additional Evidence

We sincerely thank all reviewers for their constructive feedback. Based on your suggestions, we conducted additional experiments and clarified several important aspects of CanonEdit. The main updates are summarized below.

  1. Validating the core motivation. We conducted a controlled experiment where different selection methods use exactly the same successful candidate updates. CanonEdit achieves the lowest held-out interference (0.2504), compared with max-margin (0.4105), first-valid (0.3493), and random-valid (0.3482). This directly shows why selecting among successful updates matters.

  2. Clarifying the methodological contribution. We explained the distinction between the local quadratic optimization and canonical selection across verified candidates. The new controlled experiment isolates the benefit of representative selection beyond candidate construction and verification.

  3. Strengthening evaluation reliability. We separated editing inputs from evaluation prompts and added stricter prediction metrics. At the same 97.33% pairwise editing efficacy, CanonEdit improves target Top-1 accuracy from 2.22% to 9.99% and held-out paraphrase Top-1 from 3.39% to 9.63%. We also reported independent paraphrase and protection results to make the trade-offs clearer.

  4. Examining historical retention. We clarified the difference between Hist. and All-Prior Retention and evaluated all previous edits at checkpoints 200, 400, and 600. At checkpoint 600, CanonEdit achieves higher prior-paraphrase retention (62.33% vs. 57.67%), while no-canonical performs slightly better on prior-target retention (68.83% vs. 66.33%). We also clarified the different request streams behind Tables 1 and 15.

  5. Evaluating nonlinear refinement. We examined how correction affects editing success and reliability. In the original 10K experiment, 93.90% of requests follow the closed-form path, with only 6.13% requiring re-linearization. Additional experiments show that increasing the refinement budget can improve editing success and reduce failed writes.

  6. Clarifying flexibility and baseline settings. We explained how additional editing objectives can be incorporated through constraints and weighted protection costs. We also clarified AlphaEdit's evaluation settings, target construction, and the reasons for differences across reported results.

These additions provide more direct evidence for CanonEdit's main idea: among updates that successfully write the same fact, choosing a lower-interference update can help preserve existing knowledge. They also make the evaluation protocols, trade-offs, and limitations clearer.

We sincerely appreciate the reviewers' time and thoughtful suggestions. We hope these additional results address the main concerns, and we welcome any further comments or discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.