acceptodds
Under review as a conference paper at ICLR 2027

CFPP: Conflict-Free Preservation Projection for Locate-then-Edit Model Editing

Abstract

Model editing provides an efficient way to update factual knowledge in large language models (LLMs) after deployment, without retraining the full model. A widely used locate-then-edit paradigm localizes critical MLP representations and computes a targeted parameter update. However, its preservation term can penalize edit-relevant directions when the dominant subspace of its covariance overlaps with the active edit subspace. This preservation–edit conflict becomes more severe as the number of edits grows, leading to under-editing and degraded generalization. We propose CFPP (Conflict-Free Preservation Projection), a lightweight framework that modifies the preservation covariance before computing the local closed-form update. CFPP stabilizes the edit geometry through spectral whitening and applies residual-aware soft projection with trace calibration to attenuate edit-aligned components while maintaining the overall preservation budget. CFPP is designed for locate-then-edit methods such as MEMIT, PMET, and EAMET. Experiments on GPT2-XL, Mistral-7B, and LLaMA3-8B show that CFPP yields an average relative gain of 8.2% in the reported Avg. score, with improvements reaching 31.5% in individual settings. Code is available at https://anonymous.4open.science/r/CFPP-B4H7.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Model Editing, Large Language Models, Locate-then-Edit, Massive Editing, Subspace Projection

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_CmziOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CFPP, a preservation-covariance modification for locate-then-edit methods. It attenuates preservation constraints along edit-relevant directions through whitening and adaptive soft projection, followed by trace normalization. CFPP integrates with existing editors and is evaluated across three models and three knowledge-editing benchmarks.

Critical strengths

The paper identifies a concrete geometric conflict between editing and preservation and proposes a reusable modification without additional trainable parameters. Experiments cover multiple editors, editing scales, and sequential settings, with consistent improvements in aggregate scores over the corresponding base editors.

Critical weaknesses
  1. The theoretical analysis does not establish behavioral preservation. Appendix A.2 shows that projection preserves orthogonal key components, but this does not establish preservation of the original key–value mappings or bound the resulting preservation error, such as (|\Delta K_p|_F^2). Trace normalization also provides no such guarantee. The connection between the geometric argument and actual knowledge preservation therefore remains incomplete.

  2. Evaluation of general capabilities and multi-hop reasoning is insufficient. The paper lacks general-capability benchmarks before and after editing; the generation metrics in Appendix D.6 provide limited evidence in this regard. Wiki-recent includes portability, but its coverage of multi-hop reasoning versus alias-based queries is unclear. A focused general-capability evaluation and an explicit multi-hop benchmark would better establish the practical value of the method.

  3. The comparative results require clearer interpretation. AlphaEdit underperforms MEMIT in several settings, warranting clarification of baseline configurations and evaluation protocols. Although CFPP improves Avg. in all 18 paired comparisons in Table 3, specificity decreases throughout, and some gains are marginal, such as 73.55→73.62 for EAMET on Mistral/ZsRE.

Questions for authors

The same as weaknesses.

3 comments

We thank the reviewer for the constructive feedback. We have added experiments and clarified the scope of our preservation claims.

Answer1: We agree that the projection argument in Appendix A.2 leaves open how the subsequent weight update affects the original key–value mappings. We added Appendix A.3 to quantify this effect. For fixed original preservation keys (K_p) at a single edited layer, the squared change in their output values is (|\Delta K_p|_F^2). Equation (23) gives a direction-wise upper bound on this error in terms of the edit residual, the projection gates, and the trace calibration factor. It also follows from the update expression that a fixed key’s output is unchanged if its whitened representation is orthogonal to the full edit subspace.

With the keys, residual, and gates fixed, trace calibration strengthens the remaining preservation penalties and tightens the corresponding terms in the bound. Fully gated directions retain zero penalty. Compared with the variant without calibration, enabling calibration improves specificity by 0.59 and 0.45 percentage points on CounterFact and ZsRE, respectively, and locality by 0.90 percentage points on Wiki-recent. The bound characterizes local output drift for fixed keys at one layer. We assess preservation at the model level using specificity and locality.

Answer2: To address the concern about general capabilities, we added a focused evaluation in Table 14, comparing the unedited LLaMA3-8B model with MEMIT and MEMIT-CFPP after CounterFact editing across six benchmarks. Relative to MEMIT, CFPP performs better on MRPC, SST-2, and RTE, while scoring modestly lower on MMLU, CoLA, and NLI. Overall, the added evaluation shows that CFPP retains competitive general-task performance and improves upon MEMIT on several benchmarks.

We also added a direct multi-hop evaluation on the POPULAR split of RippleEdits (Table 8). Compared with MEMIT, CFPP improves logical generalization, both compositionality criteria, and subject aliasing, indicating stronger multi-hop and compositional editing. At the same time, its preservation and relation-specificity scores remain competitive, with only small decreases relative to MEMIT. Wiki-recent portability evaluates transfer to reasoning variants and alias-based prompts; RippleEdits provides a complementary, explicit assessment of multi-hop and compositional effects. These results show that CFPP strengthens multi-hop editing while maintaining competitive preservation and specificity.

Answer3: Table 3 reports means and standard deviations over three random seeds. The reviewer’s observation that ‘AlphaEdit underperforms MEMIT in several settings’ reflects our Table 3 setup, which evaluates massive editing rather than sequential editing. Under this setup, AlphaEdit performs below MEMIT in several settings, consistent with the large-scale editing results reported in EAMET: ROBUST MASSIVE MODEL EDITING VIA EMBEDDING ALIGNMENT OPTIMIZATION, whose experimental protocol is similar to ours.

Regarding the specificity decreases across the Table 3 pairings, we added a λ sweep comparing the efficacy–specificity frontiers of MEMIT and MEMIT-CFPP on LLaMA3-8B with 10,000 CounterFact edits. At the closest matched-specificity points, efficacy is nearly identical while generalization improves. At nearby points with lower but still comparable specificity, CFPP improves both efficacy and generalization. Across much of the tested range, CFPP improves efficacy and generalization but does not reach the base method’s highest specificity. At the smaller scale of 1,000 edits, the specificity cost is mitigated, and CFPP’s specificity can even slightly exceed that of the base method.

The blue text in the newly uploaded manuscript highlights the additional content we have added.

Dear Reviewer,

Thank you again for your time and thoughtful reviews. Could you please share whether the authors’ response has addressed your main concerns, and whether any questions remain? A brief follow-up would be very helpful as I prepare the meta-review.

Please also ensure that your review reflects your current assessment in light of the authors’ response and the discussion.

Many thanks for your continued engagement.

Best regards,
Area Chair

Thank you for the clarification, and apologies for my delayed response. I appreciate the new perspective this work offers. However, the models evaluated in the experiments date from roughly two years ago, and it is unclear whether the findings still hold for current state-of-the-art models. My concerns therefore remain, and I will keep my original score.

Reviewer_7gK7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CFPP, a lightweight enhancement to locate-then-edit model editing that mitigates a geometric failure mode: overlap between the dominant preservation covariance and the active edit subspace. CFPP stabilizes preservation geometry by (i) spectral whitening of the preservation covariance, (ii) residual-aware soft projection that attenuates edit-aligned components with adaptive gating, and (iii) trace calibration to keep the overall preservation budget. Plugged into MEMIT-, PMET-, and EAMET-style updates, CFPP consistently improves efficacy and generalization across GPT2-XL, Mistral-7B, and LLaMA3-8B on CounterFact, ZsRE, and Wiki-recent.

Critical strengths
  1. Practical intervention. CFPP changes the covariance used in an existing local update, requiring no retraining.
  2. Consistent gains across methods and settings. Table 3 shows an Avg. gain in all 18 CounterFact/ZsRE pairs, and Table 4 shows gains on Wiki-recent.
Critical weaknesses
  1. Specificity falls in every pairing. In Table 3, specificity decreases in every CFPP-versus-base comparison on both CounterFact and ZsRE.
  2. No tuned-λ\lambda baseline. CFPP reduces the preservation penalty along edit directions, but the paper does not compare its efficacy–specificity frontier with a suitably tuned preservation strength λ\lambda for the base methods.
Questions for authors
  1. Do you have results from sweeping the base method’s preservation strength λ\lambda? Can you plot Efficacy against Specificity for λ\lambda versus CFPP's τ\tau? Does CFPP dominate that frontier?
  2. Table 3 reports lower specificity for every CFPP pairing on CounterFact and ZsRE. How do you interpret this result relative to the preservation claim?
3 comments

We thank the reviewer for highlighting the preservation–editing trade-off.

Answer1: We added a λ sweep for MEMIT and MEMIT-CFPP on LLaMA3-8B with 10,000 CounterFact edits (Sec. 6.1, Fig. 6). At the closest matched-specificity points (λ=10,000 for MEMIT and λ=30,000 for MEMIT-CFPP), efficacy is reasonably similar, while generalization improves from 44.02 to 45.37. CFPP also improves both metrics at nearby, nearly matched lower-specificity points. The results show better efficacy and generalization across much of the tested range at comparable specificity, alongside a specificity cost. At 1,000 edits, Table 6 shows that CFPP improves all four reported metrics on both CounterFact and ZsRE, suggesting that this cost is mitigated at smaller editing scales.

Answer2: At large editing scales, some decline in specificity or locality is expected. The soft projection may also attenuate a very small number of preservation directions that are related to editing but do not themselves conflict with it. Table 3 thus reflects a trade-off: gains in efficacy and generalization raise Avg. in every paired comparison, while specificity decreases only modestly. We characterize CFPP as mitigating the geometric conflict while retaining competitive preservation, rather than eliminating all preservation costs. At smaller editing scales, such as 1,000 edits, CFPP maintains specificity well; Table 6 shows slightly higher specificity than MEMIT on both datasets.

The blue text in the newly uploaded manuscript highlights the additional content we have added.

Dear Reviewer,

Thank you again for your time and thoughtful reviews. Could you please share whether the authors’ response has addressed your main concerns, and whether any questions remain? A brief follow-up would be very helpful as I prepare the meta-review.

Please also ensure that your review reflects your current assessment in light of the authors’ response and the discussion.

Many thanks for your continued engagement.

Best regards,
Area Chair

Thanks for the additional experiments. The λ\lambda sweep partially addresses my concern, showing improved generalization at comparable specificity in one setting. However, the evidence remains limited, and the systematic specificity decreases at larger editing scales remain a concern. The 1,000-edit results are encouraging but do not fully resolve this issue. I therefore maintain my rating, while acknowledging the additional evidence.

Reviewer_1eYKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the preservation-editing conflict in locate-then-edit knowledge editing methods. We identify a latent geometric failure mode: the primary dominant subspace of the preservation covariance overlaps with the edit-relevant subspace, whereby the preservation constraint suppresses the updates necessary for successful editing. To address this issue, we propose Conflict-Free Preservation Projection (CFPP), a lightweight framework that modifies the preservation covariance before computing closed-form parameter updates. Integrating spectral whitening, residual-aware adaptive soft projection, and trace calibration, CFPP mitigates penalties along edit-relevant directions while retaining the overall covariance budget. The framework introduces no additional trainable parameters and can be integrated into existing approaches such as MEMIT, PMET, and EAMET. We evaluate CFPP on GPT2-XL, Mistral-7B, and LLaMA3-8B using CounterFact, ZsRE, and Wiki-recent. Experiments demonstrate consistent improvements in overall editing performance across all evaluated settings, corroborated by geometric diagnostics, ablation studies, extended experiments, sequential editing evaluations, and efficiency analysis. Our core contribution lies in a new perspective for knowledge editing that emphasizes shaping the geometric structure of preservation constraints, rather than directly restricting parameter updates.

Critical strengths
  1. This work presents a valuable perspective on preservation-induced editing interference. It highlights that preservation constraints in locate-then-edit methods can block critical directions for knowledge updates. Instead of directly limiting parameter updates as prior subspace methods do, CFPP reshapes the geometry of preservation covariance. This constitutes a genuine methodological advance rather than a trivial projection variant and helps frame preservation as a structured geometric constraint to motivate future interference-aware knowledge editing research.

  2. CFPP delivers consistent empirical gains and integrates well with prior methods. The comprehensive experiments cover three model families, three benchmarks and three locate-then-edit baselines. It boosts overall editing scores in all 18 paired comparisons in Table 3, particularly for efficacy and generalization. Validated via geometric diagnostics, ablations, scalability tests, sequential editing and efficiency analysis, CFPP requires no additional trainable parameters and incurs low computational cost, making it practically attractive.

Critical weaknesses
  1. There is insufficient evidence that CFPP fundamentally improves the fidelity–editing trade-off. Although CFPP increases aggregated editing scores, the gains may come from relaxed preservation constraints rather than resolved conflicts. Across all 18 comparisons in Table 3, specificity falls while efficacy and generalization rise, matching the effect of weaker regularization. Without controlled comparisons against finely tuned baselines or Pareto-frontier validation, the claim that CFPP resolves conflicts without harming preservation is poorly justified, undermining its core argument.

  2. The proposed mechanism lacks clear component isolation and solid behavioral-level theoretical support. Table 6 ablations show spectral whitening drives most gains, with residual-aware gating adding little, and trace calibration’s individual effect unseparated. It is unclear whether simpler covariance transformations can perform similarly. Furthermore, reduced subspace overlap is not proven to causally improve editing, and the theory only characterizes the projection operator without guaranteeing preservation of unrelated predictions. These gaps disconnect the theoretical motivation, mechanism, and empirical results.

Questions for authors

Q1. Can you verify that CFPP’s trade-off improvement goes beyond weaker regularization? Specificity consistently drops in Table 3. Please compare against finely tuned baselines and provide Pareto curves or matched-specificity results to confirm better editing under comparable preservation.

Q2. Can you establish causality between subspace overlap and editing performance? Current results only show correlation, confounded by covariance spectrum, conditioning, etc. Please run controlled experiments that isolate overlap while fixing other properties.

Q3. What are the separate effects of adaptive gating and trace calibration? Adaptive gating yields small gains; please check consistency across models/seeds. Ablate trace calibration against a whitened uniform-projection baseline and clarify the scope of preservation guarantees after covariance rescaling.

2 comments

We thank the reviewer for identifying the central issue: higher aggregate scores alone do not establish a better preservation–editing trade-off. We have added analyses to address this concern and will qualify the scope of our conclusions accordingly.

Answer1: In Sec. 6.1 and Fig. 6, we sweep λ for MEMIT and MEMIT-CFPP on LLaMA3-8B with 10,000 CounterFact edits. At the closest matched-specificity operating points (λ=10,000 for MEMIT and λ=30,000 for MEMIT-CFPP), efficacy is broadly comparable, while generalization increases from 44.02 to 45.37 (+1.35 points). CFPP also improves both metrics at nearby, nearly matched lower-specificity points. At 1,000 edits, Table 6 shows higher efficacy, generalization, specificity, and Avg. for CFPP on both CounterFact and ZsRE. Our claim is therefore that CFPP improves efficacy and generalization across much of the tested range, with a specificity cost at large editing scales. This cost is mitigated at the smaller 1,000-edit scale.

Answer2: We investigate the proposed mechanism through two complementary experiments. Table 1 compares preservation covariances constructed from the same number of Wikipedia samples but exhibiting different measured levels of subspace overlap. Lower overlap is associated with larger mean principal angles and better generalization, specificity, and Avg. CFPP then directionally transforms the preservation covariance through whitening and residual-aware projection while maintaining the overall covariance trace budget. Consistent with this mechanism, Tables 2 and 10 show that CFPP reduces subspace overlap across all tested models and datasets; Table 3 reports higher Avg. in every paired setting.

Together, these two sets of results—controlled covariance comparisons and the targeted intervention implemented by CFPP—support our mechanistic interpretation: overlap between preservation and edit directions induces editing interference, and reshaping the preservation geometry mitigates it. We therefore consider the existing experiments to provide mechanistic evidence beyond simple correlation in support of the central claim of CFPP.

Answer3: Table 7 already includes the requested whitened uniform-projection baseline as the “w/o adaptive” row; we additionally added a separate “w/o calibration” ablation to isolate the contribution of trace calibration. Adaptive gating assigns projection strengths according to predicted edit energy and yields modest gains over a uniform gate matched by median strength, indicating a better efficacy–preservation balance. Trace calibration restores the stabilized covariance trace and helps retain specificity and locality, mitigating further preservation loss. At the framework level, the main experiments show consistent gains across three model families and three random seeds, supporting the robustness of the overall approach.

The blue text in the newly uploaded manuscript highlights the additional content we have added.

Dear Reviewer,

Thank you again for your time and thoughtful reviews. Could you please share whether the authors’ response has addressed your main concerns, and whether any questions remain? A brief follow-up would be very helpful as I prepare the meta-review.

Please also ensure that your review reflects your current assessment in light of the authors’ response and the discussion.

Many thanks for your continued engagement.

Best regards,
Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.