acceptodds
Under review as a conference paper at ICLR 2027

When Weight Fidelity Misleads: Concept-Erasure Safety Edits Under Diffusion-Model Quantization

Abstract

Concept-erasure edits are usually evaluated before quantization, although deployed models often run at lower numerical precision. We ask whether preservation of an edit in weight space predicts preservation of its function and behavior. On Stable Diffusion 1.5, we separately measure the deployed weight difference, the internal activation difference induced by the edit, and generated outputs. Under NF4 weight quantization, UCE's deployed weight difference has cosine similarity 0.568 with the original edit, while its activation-level effect remains aligned at 0.918. For the nudity edit, projecting the differential weight error onto edit-relevant conditioning inputs reduces its relative magnitude from 1.29 in weight space to 0.44. ESD-u also exhibits substantial distortion of the deployed weight difference while retaining high average activation alignment, extending the parameter–function decoupling to a different editing surface. We next quantize the text-conditioning inputs to selected cross-attention projections. At four activation bits, NudeNet efficacy falls by 0.264–0.290 for UCE and TIME, and CLIP measurements show the same direction, with all four paired intervals excluding zero. The INT8 weights remain identical across activation precisions. An exploratory 48-prompt, six-seed isolation with INT8 weights fixed attributes this change to activation rounding. Raw edited-image nudity scores also rise, matched-magnitude Gaussian controls do not reproduce the weakening, and Dreamlike Photoreal 2.0 reproduces the A4 effect with paired intervals excluding zero. On 50 unrelated prompts, CLIPScore declines similarly for base and edited models, providing no evidence of selective edited-model collapse under this probe. Our evidence is limited to one safety target, related SD-1.5 checkpoints, and simulated quantization of selected projections. Nevertheless, it shows that weight-space fidelity is insufficient: safety edits should be evaluated under the numerical configuration in which they will be deployed.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
concept erasure, machine unlearning, post-training quantization, diffusion models, text-to-image generation, AI safety, activation quantization, model compression

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_JFY1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how quantization affects the reliability of concept erasure in diffusion models. The authors examine the relationship between parameter fidelity, activation-level functional fidelity, and behavioral erasure effectiveness. Experiments on Stable Diffusion 1.5 show that weight quantization can substantially distort parameter updates while largely preserving their functional effects. In contrast, low-bit activation quantization significantly weakens nudity erasure for UCE and TIME, while other concepts remain largely unaffected. The paper highlights the importance of evaluating concept erasure under deployment-time numerical configurations.

Critical strengths
  1. The research problem is Meaningful. The paper investigates an important yet underexplored issue: whether concept erasure remains effective after model quantization, a common practice in deployment.

  2. The multi-level analysis is convincing. The distinction between parameter fidelity, functional fidelity, and behavioral effectiveness provides useful insights into the limitations of weight-space metrics for evaluating erasure preservation.

  3. The design of experiments is reasonable. The paper includes multiple quantization settings, matched-noise controls, and additional checkpoints, providing empirical support for its main observations.

Critical weaknesses
  1. Lack of a unified explanation for the observed phenomena. The paper reports substantially different effects of activation quantization across concepts: nudity erasure degrades significantly, whereas Van Gogh and church erasure remain largely preserved or even improve. However, the paper does not provide a convincing explanation for this discrepancy or establish a unified understanding of when and why quantization compromises concept erasure. The connection between quantization errors and behavioral degradation also remains insufficiently explained. This may affect the evaluation of the paper’s theoretical contributions.

  2. The generalizability is limited. The main findings are primarily established on SD 1.5 with UCE and TIME, which are closely related editing approaches. The results for ESD-x are less conclusive, and the behavior on other architectures and concept-erasure methods remains unclear. Therefore, it is difficult to tell whether the phenomenon in this work is generalizable.

Questions for authors
  1. Could the authors further clarify why activation quantization significantly weakens nudity erasure while Van Gogh and church erasure remain largely unaffected or even improve?

  2. Following Q1, is there a unified explanation connecting quantization errors to these different behavioral outcomes?

  3. Could the authors further discuss whether the observed phenomena are specific to UCE and TIME on SD 1.5, or whether they can generalize to other concept-erasure methods and model architectures? Additional evidence would help strengthen the generality of the conclusions.

I'm open to raising my score if the authors can address these concerns.

3 comments
AuthorsOpenReview ↗

Response to JFY1: Causal Interventions, Concept Dependence, and Generalization

Thank you for recognizing the value of separating parameter fidelity, functional effects and behavioral erasure. Your request for a connection between quantization error and behavior motivated the new interventions and concept analysis in Section 6.

Q1-Q2: What connects activation quantization to behavioral weakening, and why do concepts differ?

We now identify a specific failure pathway by intervention. With deployed weights fixed, rounding only the BOS token's cross-attention value inputs weakens nudity erasure for UCE, TIME and a published ESD checkpoint. NudeNet efficacy changes are -0.243 [-0.454, -0.030], -0.224 [-0.434, -0.036] and -0.167 [-0.255, -0.088], respectively. CLIP agrees, and edited-model nudity scores rise, so the finding is not solely a normalization effect.

The complementary intervention preserves those inputs under scoped W8A4. It improves normalized efficacy for all three erasers on both detectors and directly outperforms preserving EOS: for example, UCE +0.201 [+0.052, +0.352] on NudeNet. Preservation is not established as equivalent to W8A16, and NudeNet absolute-suppression gains remain unresolved. Figure 2 and Table 3 report the interventions.

The concept analysis distinguishes outcomes that the original normalized metric obscured:

  • Van Gogh: base and edited target scores fall by similar fractions, approximately 29% and 26%. Efficacy changes by -0.012 [-0.094, +0.065], consistent with proportional attenuation rather than unchanged generation.
  • Church: the edited target score falls more strongly, while edited-model utility also deteriorates. The efficacy increase therefore accompanies generation degradation.
  • Nudity and violence: edited target scores rise, by +0.105 [+0.023, +0.196] on NudeNet and +0.083 [+0.011, +0.156] on Q16, respectively. These increases contribute to declining efficacy.

Writing efficacy as η=1−E/B\eta = 1-E/B, with base and edited target scores B,EB,E and changes δB,δE\delta B,\delta E, its change is exactly

Δη=EδB−BδEB(B+δB).\Delta\eta = \frac{E\delta B-B\delta E}{B(B+\delta B)}.

This separates changes in the base score from changes in the edited score. It explains the metric's direction, not why the network produces those changes.

The edit-strength titration provides an additional distinction: for violence, the difference between the fully edited and base models' responses is resolved, +0.142 [+0.041, +0.254]. However, the complete prespecified three-contrast account of edit counteraction is not established; the corresponding interaction for nudity is unresolved on NudeNet. We therefore provide a common intervention pathway and differentiated behavioral evidence, without claiming a universal predictor of concept direction. See Figure 3 and Tables 15-16.

Q3: How far do the results generalize?

The published ESD checkpoint adds a fine-tuned eraser beyond UCE/TIME, and violence adds a second safety target with UCE. Same-GPU violence tests show BOS-only weakening, -0.175 [-0.334, -0.047], and an advantage of BOS over EOS preservation, +0.119 [+0.020, +0.230], on Q16.

We also evaluated FLUX.1-dev with an EraseAnything LoRA under actual SVDQuant INT4 execution and MixDQ's W8A8 simulation on SDXL-turbo. Neither shows statistically resolved weakening in the tested configuration. These extend coverage while bounding the claim: they do not establish transfer of the SD-1.x mechanism or equivalence to full precision. The FLUX adapter executes in a 16-bit branch, and implementation differences are not isolated. Tables 17-18 report the results and limitations.

All intervals are 95% paired prompt-bootstrap intervals. The general response details our correction of the submitted ESD E1/E4 base-checkpoint mismatch.

We hope the intervention evidence and concept-specific analysis address the central concerns behind your assessment. We would welcome your reassessment and clarification of any remaining substantive concerns.

Thank you for your response and additional experiments. I appreciate the authors' efforts to address my concerns. The new causal interventions and concept-specific analyses provide stronger empirical evidence and help clarify the observed phenomena. However, the underlying mechanism explaining why different concepts respond differently to quantization remains insufficiently understood. While the rebuttal strengthens the paper, my major concern is not fully addressed so I will maintain my original rating of 2.

AuthorsOpenReview ↗

Clarification of the Mechanistic Finding and Its Remaining Limit

Thank you for reassessing the revision and acknowledging the stronger intervention evidence and clearer concept-specific analysis. We agree that the network-level explanation for why the same perturbation produces different target-score directions remains incomplete. The revision distinguishes two questions. First, which computation can cause erasure weakening? The fixed-weight rounding and complementary preservation interventions localize a sufficient failure pathway to the BOS value inputs within the tested SD-1.x setting. Second, why does perturbing that pathway increase some concept scores and decrease others? The titration and decomposition characterize those outcomes but do not fully answer this second question. We therefore do not present the decomposition as a complete mechanistic explanation. Its narrower contribution is to show that apparently preserved or improved normalized efficacy can accompany substantial behavioral changes, including attenuation of the base concept or generation degradation. We understand that this remaining explanatory gap continues to affect your assessment. We respectfully maintain that experimentally identifying and intervening on an erasure failure pathway, alongside demonstrating the limitations of weight-fidelity diagnostics, constitutes a useful contribution within the stated scope. We have made the unresolved concept-direction question explicit in the revision. Thank you again for your careful review and engagement.

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

This paper studies how low-bit quantization affects concept erasure in diffusion models. The authors compare weight changes, activation differences and generation behavior before and after quantization. The results show that weight similarity can be low while activation effects remain largely preserved. However, 4-bit activation quantization weakens nudity erasure. The authors also conduct Gaussian noise controls and experiments with different editing methods and checkpoints.

Critical strengths
  1. Interesting observations. The paper shows that weight similarity does not always reflect functional preservation and low-bit activation quantization can weaken concept erasure.

  2. Detailed experimental analysis. The authors examine weight differences, activation changes and generation behavior separately. Also this paper includes different experiment configurations of quantization settings, checkpoints and Gaussian noise controls.

Critical weaknesses
  1. Limited novelty. The paper reports some interesting findings without proposing a new method or general framework. Quantization-induced unlearning failures have been studied in prior work (Zhang et al., ICLR 2025) and the paper doesn’t clearly establish the differences in mechanisms.

  2. Limited generalization. The paper includes complete experiments focusing on nudity erasure and SD1.5 but lacks evaluations on modern diffusion architectures or other model adaptations like LoRA.

  3. Insufficient mechanism analysis. Although the authors observe that activation quantization weakens erasure, they do not clearly identify which components or operations cause this behavior.

Questions for authors
  1. How do these findings differ from previous studies on quantization-induced failures in model unlearning? Do similar effects occur in more general adaptations, such as LoRA or other customized methods?
  2. Do the results generalize to modern diffusion architectures, such as DiT and practical quantization pipelines?
  3. Can the authors further explain which activation components or attention operations cause the observed weakening of concept erasure?
3 comments
AuthorsOpenReview ↗

Response to wtnB: Novelty, Responsible Operations, and Practical Quantization

Thank you for asking us to distinguish the contribution from prior unlearning work, identify the responsible operations, and evaluate modern architectures and practical pipelines. The revision adds evidence addressing each point.

1. Novelty relative to Zhang et al. and diffusion quantization work.

Zhang et al.'s weight-quantization account concerns small unlearning updates mapping back into the original weight bins. Our intervention comparisons use identical deployed weights and change only activation precision. Rounding only the BOS token's cross-attention value inputs weakens nudity erasure: for UCE, -0.243 [-0.454, -0.030] on NudeNet. An exploratory full-precision-weight control also weakens under activation rounding, so INT8 weight quantization is not required in that setting. The two mechanisms may coexist, but weight-update collapse cannot explain these fixed-weight contrasts.

We explicitly credit MixDQ for identifying BOS sensitivity and preserving BOS K/V in higher precision. We do not claim either observation as new. Our additional contribution is erasure-specific intervention evidence: BOS value-input rounding induces weakening under per-token quantization for UCE, TIME and ESD, while preserving those inputs improves normalized efficacy and directly outperforms EOS preservation. This complements the original finding that weight-space fidelity alone can misrepresent erasure preservation.

2. Which operation causes the observed weakening?

The interventions localize a sufficient pathway to the BOS value inputs. Value-only and BOS-only rounding weaken erasure across the three nudity erasers. Key-only rounding was tested only for UCE and does not resolve an effect; that comparison also has a disclosed GPU-type confound. We therefore do not claim that key perturbations are generally irrelevant.

At fixed queries and keys, the BOS value perturbation changes an attention-head output by its projected error weighted by attention to BOS. The rounding and preservation experiments establish the behavioral effect; projected-error measurements characterize the pathway but are not treated as independently established mediation.

Preservation does not establish equivalence to W8A16, and its NudeNet absolute-suppression gains remain unresolved. Figure 2 and Table 3 report the principal contrasts.

3. Modern architectures, LoRA and practical quantization.

We now include two deployment evaluations:

  • FLUX.1-dev + EraseAnything LoRA under Nunchaku SVDQuant INT4: actual low-bit execution shows no statistically resolved efficacy change, -0.017 [-0.236, +0.242] on NudeNet and +0.016 [-0.026, +0.062] on CLIP. The execution audit documents W4A4 transformer-block linears and a 16-bit LoRA branch. This tests a supported deployment of a transformer-based LoRA eraser; it does not test quantizing the adapter's own weights. Implementation differences and the wide NudeNet interval limit interpretation.
  • SDXL-turbo + our UCE-style edit under official MixDQ W8A8 simulation: no resolved efficacy change, +0.067 [-0.019, +0.155] on NudeNet and -0.010 [-0.030, +0.010] on CLIP. The FP16 edit is weak. Disabling BOS-aware precision degrades generation globally, preventing an erasure-specific interpretation of that ablation.

We report these null results as boundaries, not as evidence of universal robustness or confirmation of the SD-1.x mechanism. Tables 17-18 and Appendix A distinguish simulation, kernel execution and adapter precision.

We additionally tested violence with UCE: same-GPU BOS-only rounding weakens efficacy by -0.175 [-0.334, -0.047] on Q16, and BOS preservation outperforms EOS preservation by +0.119 [+0.020, +0.230].

All intervals are 95% paired prompt-bootstrap intervals. The ESD E1/E4 base-checkpoint correction is disclosed in the general response and Appendix D.

We hope these additions clarify both the distinct scientific contribution and its deployment scope, and would appreciate your reassessment in light of this evidence.

Thank you for the detailed response and additional experiments. The fixed-weight comparisons clarify the distinction from prior weight-quantization work, and the BOS value-input interventions provide a more concrete explanation of the observed weakening.

However, as acknowledged in the response, BOS sensitivity and preserving BOS inputs in higher precision have already been studied, which limits the novelty of this aspect of the contribution. The FLUX/LoRA and MixDQ experiments broaden the evaluation, but do not establish whether the identified failure pathway generalizes to other architectures and practical quantization settings. Therefore, I will maintain my rating.

Thank you for reconsidering the paper and acknowledging that the fixed-weight comparisons distinguish our findings from prior weight-quantization work and that the interventions provide a more concrete explanation.

We agree that BOS sensitivity and high-precision BOS preservation are not new, and we explicitly credit MixDQ for these findings. Our novelty claim concerns what these numerical effects imply for concept erasure.

The revised experiments establish that, with the deployed weights unchanged, perturbing the BOS value inputs can increase edited-model target expression and weaken erasure across three nudity erasers. Conversely, preserving these inputs improves normalized erasure efficacy and directly outperforms EOS preservation. Thus, the contribution is evidence connecting a specific numerical intervention to the survival of an intentional model edit, rather than the discovery of a previously unknown sensitive token or a new general-purpose quantization method.

This also complements the paper's original measurement result: weight-space fidelity can remain identical across deployment configurations while behavioral erasure changes. The intervention experiments explain one concrete way this diagnostic limitation arises. Prior evidence about BOS sensitivity and generation quality alone does not establish these erasure outcomes.

We agree that the FLUX/LoRA and MixDQ evaluations do not demonstrate transfer of the identified pathway. We report their unresolved weakening contrasts as boundaries of the evidence, not as mechanism replications or proof of robustness. The pathway claim remains scoped to the tested SD-1.x setting; the broader evaluations address the behavioral question under additional deployment configurations.

We appreciate that the significance of this erasure-specific contribution remains a point of disagreement. We hope the distinction above helps clarify what is established beyond prior BOS sensitivity work, while preserving the limitations you identify.

Reviewer_XT8kOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how quantization after the concept erasure can affect the effectiveness of concept erasure in diffusion models, showing that erasure can weaken after quantization but not always. The authors use experiments and analysis to show the mismatch between weight changes, activation-level effects, and behavioral outcomes, and state that it is insufficient to use weight similarity alone to check erasure's effectiveness after quantization.

Critical strengths

The paper provides relatively adequate experiments on the nudity-erasure case, showing that changes in weight-space similarity alone cannot fully explain the effectiveness of concept erasure after quantization. Corresponding evaluations and Gaussian noise control experiments can support this statement.

Critical weaknesses

The results of this paper are mainly validated on the nudity task, while the Van Gogh and Church cases do not exhibit the same degradation in erasure effectiveness. The statements are also limited in SD1.5 architecture diffusion models, and UCE and TIME editing methods. Thus, the statement proposed in this paper can hardly be generalized.

The authors do not explain the reasons for the failure in the nudity erasure case. Moreover, the mathematical analysis cannot fully support why activation quantization weakens the effectiveness of nudity erasure. Therefore, the paper lacks a clear mechanistic understanding of the observed phenomenon.

Questions for authors

Could the authors provide explanations about why failure in erasure effectiveness is observed only for nudity cases but not for the other concepts cases. Moreover, can the authors provide a concrete mechanism explaining why activation quantization causes this failure?

3 comments
AuthorsOpenReview ↗

Response to XT8k: Failure Mechanism and Expanded Evaluation Scope

Thank you for identifying the missing mechanism and limited generality. The revision adds targeted interventions, a concept-direction analysis and broader deployment evaluations.

What causes the nudity-erasure failure?

With deployed weights fixed, rounding only the BOS token's cross-attention value inputs is sufficient to weaken nudity erasure for UCE, TIME and a published ESD checkpoint on its correct SD-1.4 base. NudeNet efficacy changes are -0.243 [-0.454, -0.030], -0.224 [-0.434, -0.036] and -0.167 [-0.255, -0.088], respectively. CLIP agrees, and raw edited-model nudity scores rise.

The complementary intervention preserves those inputs under scoped W8A4. Normalized efficacy improves for all three erasers on both detectors, and BOS preservation directly outperforms EOS preservation. This localizes a failure pathway through behavioral interventions rather than relying on weight similarity or mathematical identities alone. It does not establish complete recovery or resolved improvement in NudeNet absolute suppression. See Section 6, Figure 2 and Table 3.

Why does the outcome differ across concepts?

The new analysis shows that the benign concepts are also affected:

  • Van Gogh: base and edited scores fall by similar fractions. Approximately stable normalized efficacy is consistent with proportional attenuation, not unchanged behavior.
  • Church: edited target scores fall more strongly, while utility deteriorates. Improved normalized efficacy therefore accompanies generation degradation.
  • Nudity and violence: edited-model target scores rise and efficacy declines.

An exact decomposition separates the contributions of changing base and edited scores. The edit-strength titration also resolves a differential response between the fully edited and base violence models, +0.142 [+0.041, +0.254] on Q16, although the full prespecified account of edit counteraction remains unestablished.

These findings explain how the different efficacy directions arise from observable behavioral changes. They do not yet explain why the network produces each concept's score changes. See Figure 3 and Tables 15-16.

How broad is the evidence?

ESD adds a fine-tuned eraser beyond UCE/TIME, and violence adds a second safety target with UCE. We also tested FLUX.1-dev with an EraseAnything LoRA under actual SVDQuant INT4 execution, and SDXL-turbo under MixDQ's official W8A8 simulation. Neither shows statistically resolved weakening in the tested configuration. These results extend the evaluation beyond SD-1.5 while making clear that the failure is conditional, not universal. The FLUX adapter itself executes in a 16-bit branch, and its deployment comparison does not isolate quantization from implementation differences.

All intervals are 95% paired prompt-bootstrap intervals. The general response documents the corrected ESD E1/E4 base mismatch.

We hope the localized intervention evidence and expanded evaluation address your principal concerns, and would welcome your reassessment of the revised contribution.

Thank you for the detailed responses and additional experiments. The new BOS interventions provide stronger evidence and experiments also show a better generalizability.

However, my main concerns remain partially unresolved. While the analysis clarifies the different efficacy trends, it does not fully explain why quantization weakens erasure for nudity and violence but produces different outcomes for Van Gogh and Church. Furthermore, the additional evaluations do not establish whether the identified failure pathway extends to other architectures and quantization settings. Therefore, it remains difficult for me to determine when or whether these findings generalize to other concepts and datasets.

Overall, I appreciate the improvements, but these are still unresolved concerns. I will maintain my rating of 1.

Thank you for reconsidering the revision and acknowledging that the BOS interventions strengthen the evidence and that the additional experiments broaden the evaluation. We agree that predicting concept-specific outcomes and establishing transfer of the same pathway beyond SD-1.x remain open questions.

We would like to clarify what the revised evidence establishes despite these limitations.

The failure pathway is experimentally localized within the tested setting. With deployed weights fixed, rounding the BOS value inputs is sufficient to weaken nudity erasure across UCE, TIME and a published ESD checkpoint. The complementary preservation intervention improves normalized efficacy and directly outperforms EOS preservation. Violence with UCE provides a second safety target with the same intervention pattern. These results identify a computation through which activation precision changes erasure; they do not establish why every concept responds in a particular direction.

The concept analysis distinguishes materially different behavioral outcomes. Approximately stable Van Gogh efficacy accompanies falling base and edited scores, while improved church efficacy accompanies generation degradation. Nudity and violence instead show rising edited-model target scores. This analysis does not predict unseen concepts, but it clarifies why stable or improved normalized efficacy cannot by itself establish preserved behavior.

The broader evaluations delimit the finding. The FLUX/LoRA INT4 and MixDQ evaluations show no statistically resolved weakening in their tested configurations. We do not treat these as replications of the BOS pathway or evidence of equivalence. They extend the deployment evaluation while leaving transfer of the mechanism unresolved.

We therefore agree that the paper does not yet provide a predictive account across concepts, datasets and architectures. Its demonstrated contribution is the separation of weight fidelity from behavioral erasure preservation, an experimentally localized activation-induced failure pathway, and a tested precision-preservation intervention within the stated scope. We believe these findings provide a useful basis for evaluating erased models in their deployed numerical configuration, while the remaining generalization questions define the limits of the current evidence.

Thank you again for your careful consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions: Mechanism, Concept Dependence, and Deployment Evaluation

We thank the reviewers for their constructive feedback. Their shared concerns motivated three substantial additions: interventions identifying a failure pathway, an analysis distinguishing concept-dependent outcomes, and evaluations beyond SD-1.x using practical quantization pipelines. We have updated both the manuscript and the supplementary package.

  1. From an observed failure to an experimentally localized pathway.

With deployed weights fixed, rounding only the BOS token's cross-attention value inputs is sufficient to weaken nudity erasure for UCE, TIME and a published ESD checkpoint on its correct SD-1.4 base. NudeNet normalized-efficacy changes are -0.243 [-0.454, -0.030], -0.224 [-0.434, -0.036] and -0.167 [-0.255, -0.088], respectively. CLIP agrees, and raw edited-model nudity scores also rise.

The complementary intervention preserves these inputs under scoped W8A4, which rounds the cross-attention K/V inputs to four bits with INT8 projection weights. Preservation improves normalized efficacy for all three erasers on both detectors and directly outperforms preserving EOS. These interventions identify a localized failure pathway, while complete recovery to W8A16 and gains in NudeNet absolute suppression remain unestablished. See Section 6, Figure 2 and Table 3.

  1. Explaining the different measured outcomes across concepts.

The new edit-strength titration and raw-score decomposition distinguish three patterns:

  • Nudity and violence: edited-model target scores rise, contributing to reduced efficacy.
  • Van Gogh: base and edited scores fall by similar fractions, consistent with proportional attenuation and approximately stable normalized efficacy.
  • Church: the edited target score falls more strongly, while generation utility deteriorates; increased normalized efficacy therefore accompanies degradation.

For violence, the difference between the fully edited and base models' responses is also resolved: +0.142 [+0.041, +0.254] on Q16. The complete prespecified account of edit counteraction remains unresolved.

This analysis explains how the efficacy directions arise from observable changes in base and edited scores. It does not claim to explain why the network produces each concept's score changes. See Figure 3 and Tables 15-16.

  1. Broader evaluation, including a transformer, a LoRA eraser and real INT4 execution.

We added violence with UCE and evaluated FLUX.1-dev with a released EraseAnything LoRA under actual Nunchaku SVDQuant INT4 execution. The FLUX comparison shows no statistically resolved weakening: -0.017 [-0.236, +0.242] on NudeNet and +0.016 [-0.026, +0.062] on CLIP. The adapter executes in a 16-bit branch, and the comparison does not isolate quantization from implementation differences.

We also ran MixDQ's official W8A8 simulation on SDXL-turbo with our UCE-style edit, finding no resolved weakening in that configuration. These evaluations address the requested breadth while establishing boundaries: the SD-1.x failure is not claimed to occur universally, and neither deployment test establishes equivalence to full precision or tests the same BOS pathway. See Tables 17-18 and Appendix A.

Correction to the submitted ESD experiments.

The submitted E1/E4 experiments paired the published SD-1.4 ESD checkpoint with SD-1.5. We reran both on the correct SD-1.4 base. The A4 boundary interpretation remains, but the CLIP explanation changes to an increasing edited score. NF4 now shows reduced NudeNet absolute suppression, while normalized-efficacy changes remain unresolved. Section 5.1 and Appendix D document the correction. UCE, TIME and locally trained ESD experiments are unaffected.

Revised manuscript and supplementary package.

The revised manuscript updates the abstract, adds the intervention and concept analyses, reports the broader deployment evaluations, and clarifies the limits of each conclusion.

The updated anonymous supplementary package includes the associated code, configurations, retained outputs, corrected ESD results, and reproduction instructions. Its README and change log identify the additions and corrections, distinguish exploratory and confirmation stages, and document data reuse, hardware differences and execution details.

All intervals above are 95% paired prompt-bootstrap intervals. The revised contribution combines the original evidence that weight fidelity can misrepresent erasure preservation with interventions showing how activation precision can weaken erasure despite identical deployed weights. We hope these additions address the reviewers' central concerns and provide a stronger basis for reassessing the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.