Abstract

Pre-trained Vision Transformers (ViTs) are increasingly deployed for medical image classification. However, correcting their inevitable failure cases in dynamic clinical scenarios poses a critical challenge. Conventional fine-tuning approaches inherently suffer from catastrophic forgetting, severely degrading previously acquired diagnostic capabilities. Such instability fundamentally compromises clinical safety. Addressing this vulnerability requires an active, controllable, and reliable intervention mechanism that is both theoretically grounded and inherently interpretable. To this end, we propose X-Edit (eXact, eXplicit, and eXplainable Editing), an efficient null-space model editing framework. X-Edit transitions the editing process from iterative gradient-based optimization to a theoretically grounded, closed-form solution. Specifically, we first explicitly localize the influential layers via causal tracing governing the erroneous prediction. Subsequently, we construct an orthogonal null-space projection matrix from a curated anchor set. By geometrically constraining the exact parameter update strictly within this null space, we provide mathematical guarantees that the intervention rectifies targeted errors without perturbing established diagnostic representations. Extensive evaluations on six medical imaging benchmarks demonstrate that X-Edit comprehensively suppresses catastrophic forgetting while achieving superior edit success rates. Our code is available at \url{https://github.com/HenryLau7/X-Edit}.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1503_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/HenryLau7/X-Edit

Link to the Dataset(s)

MedMNIST: https://medmnist.com/ LiFS: https://zmic.org.cn/care_2026/track_liver/

BibTex

@InProceedings{LiuYua_XEdit_MICCAI2026,
        author = { Liu, Yuanye AND Zhou, Siyuan AND Zhang, Ke AND Li, Lei AND Chen, Wei AND Zhuang, Xiahai},
        title = { { X-Edit: Exact, Explicit, and Explainable Null-Space Editing for Medical Vision Transformers } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16882},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes X-Edit, a null-space constrained model editing framework for ViT in medical image classification. It identifies influential MLP layers via causal tracing, constructs an orthogonal projection onto the input null space derived from anchor samples to preserve prior behavior, and applies a closed-form parameter update that aims to fix misclassified “edit” cases while preventing catastrophic forgetting. Experiments on six medical imaging benchmarks show high edit success with markedly reduced degradation on held-out test sets and lower edit-time than iterative editing baselines.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper addresses an important issue: how to correct failure cases in deployed medical ViTs without catastrophic forgetting. This is meaningful for me.

    The overall method is more principled than simple fine-tuning and more interpretable than purely heuristic editing.

    On the reported benchmarks, X-Edit achieves strong edit success rates while substantially reducing test accuracy drop relative to FT/EWC/LWE.

    The anonymous repository is well organized, which further strengthens the paper’s completeness and reproducibility, and makes the work a potentially useful baseline for follow-up research.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The mathematical guarantee is notable narrower than the paper’s framing suggests.

    • The method starts from an anchor-constrained editing objective, but the actual derivation only guarantees invariance of the edited layer’s linear mapping on the finite anchor-induced key subspace, which is, (W+\Delta WP)K_0 = WK_0.However, it is not equivalent to preserving the full model behavior on the true test or deployment distribution. Likewise, the closed-form solution is exact only for the surrogate quadratic objective used in the derivation, not for the original nonlinear model-editing problem or for downstream clinical risk. I therefore believe the claims around “exact,” “safe,” “trustworthy,” and “lifelong clinical deployment” are overstated and should be substantially toned down.

    A minor concern is that, include other continual learning and editing baselines would be helpful to position the work.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper is technically interesting and reasonably novel, and the overall experimental package is quite complete. I especially appreciate that the work is presented as a coherent tool/framework rather than an isolated idea, and the anonymous repository suggests good reproducibility and makes the paper a potentially useful baseline for future work. My main reservation is that the paper’s mathematical framing appears stronger than what is actually established: the guarantee seems limited to anchor-level preservation under a surrogate constrained objective, while some of the writing suggests a broader notion of exactness, safety, and trustworthy clinical deployment.

  • Reviewer confidence

    Somewhat confident (2)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper introduces X-Edit, a null-space model editing framework designed to rectify errors in medical Vision Transformers without catastrophic forgetting. The method combines causal tracing for identifying influential layers with a closed-form null-space parameter update that provides mathematical guarantees against performance degradation. The framework enables exact, explicit, and explainable post-hoc interventions for correcting model failures in dynamic clinical environments.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The paper addresses a clinically critical problem: correcting model errors without catastrophic forgetting, which is essential for safe deployment of AI in healthcare.
    • The theoretical grounding is strong, with mathematical guarantees provided through null-space projection constraints.
    • The closed-form solution for parameter updates is computationally efficient compared to iterative gradient-based methods, making it practical for real-world deployment.
    • The integration of causal tracing with null-space editing provides an interpretable and controllable intervention mechanism.
    • Comprehensive evaluation across six diverse medical imaging benchmarks demonstrates robust generalization.
    • The proposed DPE (Drop Per Edit) metric effectively quantifies the editing trade-off between correction success and collateral damage.
    • The authors provide anonymized code access, supporting reproducibility.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • LIMITED THEORETICAL NOVELTY: The null-space editing approach is adapted from recent work in LLM editing (AlphaEdit, MEND). While the application to medical ViTs is new, the core technical contribution is incremental.
    • INSUFFICIENT ANALYSIS OF EDIT SCOPE: The paper does not thoroughly analyze how the number of edited samples affects performance. Real clinical scenarios may require editing on hundreds of cases, not just the small edit sets reported.
    • LACK OF CROSS-DOMAIN EVALUATION: All experiments are conducted within the same domain as the training data. The framework’s robustness to domain shifts (e.g., different hospitals, scanners) is not evaluated.
    • MISSING ABLATION ON CAUSAL TRACING: The paper does not ablate the causal tracing component to demonstrate its necessity compared to simply editing fixed layers across all samples.
    • LIMITED COMPARISON WITH SOTA: While LWE is included as a baseline, other recent model editing methods for vision models are not compared.
    • NO ANALYSIS OF EDIT ACCUMULATION: The sequential editing capability is claimed but not thoroughly validated with long edit sequences (e.g., 100+ sequential edits).
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    This is a solid contribution to the field of trustworthy medical AI. To further strengthen the paper, I suggest the authors: (1) Include ablation studies demonstrating the necessity of causal tracing versus fixed-layer editing; (2) Evaluate the framework’s robustness under domain shifts; (3) Analyze performance with larger edit sets to better simulate real-world clinical scenarios. These additions would provide valuable insights into the practical deployment of X-Edit.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    INNOVATIVENESS (Score: 3/5): The application of null-space editing to medical ViTs is novel, though the core technique is adapted from LLM editing literature. The integration with causal tracing provides a meaningful contribution to the model editing toolbox for medical imaging.

    TECHNICAL CORRECTNESS (Score: 4/5): The mathematical formulation is sound, and the closed-form solution is well-derived. The null-space constraint provides a principled way to prevent catastrophic forgetting. The experimental design is appropriate for the problem.

    EXPERIMENTAL VALIDATION (Score: 4/5): The comprehensive evaluation across six datasets demonstrates robust generalization. The comparison with multiple baselines (including EWC, FineTune, and LWE) provides good context. The proposed DPE metric is a useful contribution for quantifying editing trade-offs.

    CLINICAL IMPACT (Score: 4/5): The paper addresses a critical clinical need for safe model updating in dynamic environments. The efficiency of the closed-form solution makes it practical for real-world deployment. The mathematical interpretability aligns well with clinical AI requirements.

    PRESENTATION (Score: 4/5): The paper is well-organized and clearly written. The figures effectively illustrate the framework, and the experimental results are presented comprehensively.

    Overall, this is a solid methodological contribution with strong experimental validation and clear clinical relevance. While the core technique is not entirely novel (being adapted from LLM editing), the application to medical ViTs and the comprehensive evaluation make it a worthwhile addition to the MICCAI community. The paper would benefit from additional analysis of edit accumulation and cross-domain robustness, but these are not critical flaws.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    This paper presents X-Edit, a post-hoc model editing framework for medical Vision Transformers that aims to correct newly observed failure cases without causing catastrophic forgetting. The key idea is to combine causal tracing to identify influential layers with a null-space constrained closed-form update, so that edits are applied only in directions that preserve anchor representations. The paper positions this as an exact, explicit, and explainable editing strategy for reliable clinical deployment, and validates it across multiple medical imaging benchmarks.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    A major strength of the paper is that it addresses a highly relevant problem for medical AI: how to fix model errors after deployment without degrading previously learned diagnostic capability. This is an important and practical setting in clinical environments, where naive fine-tuning can be unsafe.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The main weakness is that the paper’s novelty seems to come more from adapting and integrating existing ideas than from introducing a fundamentally new editing principle. The use of null-space constrained editing is closely related to prior work such as AlphaEdit and related transformer editing methods, while causal tracing and ViT editing are also already explored in prior literature. Therefore, the paper’s novelty is meaningful at the application/integration level, but the conceptual leap may be somewhat incremental.

    Another weakness is that the evaluation, while broad across datasets, is still limited in terms of clinical realism and scale. Several experiments involve a very small number of edit samples, such as only 3 edited cases for OrganAMNIST and 8 for LiFS-subtask2, which makes it harder to judge robustness under larger-scale or repeated editing scenarios.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    My recommendation is based on the fact that the paper tackles an important and clinically relevant problem, and proposes a method that is technically clean, efficient, and empirically promising. In particular, I found the combination of causal tracing and null-space constrained closed-form editing to be well motivated, and the results suggest that the method achieves a better trade-off between edit success and preservation of existing performance than the included baselines.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

## Novelty

R3 & R4: The conceptual novelty feels incremental compared to LLM editing methods.

R: Thanks. We agree that X-Edit is inspired by recent null space editing methods in LLMs. However, the contribution of our work is to reformulate this principle for medical ViTs, where the edited knowledge is visual diagnostic behavior rather than textual factual associations. This requires a different localization mechanism, a different representation preservation constraint, and a clinically motivated evaluation protocol. We will clarify in the camera-ready.

## Methods

R3 & R4: Raised concerns about the small number of edit samples in some datasets and requested validation for large-scale, sequential editing.

R: We thank the reviewer for this valuable comment. The relatively small edit sets in some datasets come from our protocol: D_edit consists of naturally misclassified samples from a disjoint validation split, reflecting the intended post-hoc correction setting. Therefore, the edit size is determined by the actual failure cases of the base model. We also note that X-Edit is designed to support sequential editing. In Eq. (4) and Eq. (5), previously edited keys are explicitly included to reduce interference with past edits. Due to the limited space, we focused on the main multi-dataset reliability comparison and did not report extensive large-scale sequential curves. We will clarify this design choice and limitation in the revision.

## Writing

R2: The mathematical guarantees apply strictly to the finite anchor-induced subspace and surrogate quadratic objective, recommending toning down absolute claims like exact, safe, and lifelong.

R: We thank the reviewer for this important clarification. We agree that the mathematical guarantee of X-Edit should be interpreted within its formal scope, namely the finite anchor-induced null space and the surrogate quadratic objective. Specifically, the projection preserves anchor representations, while the closed-form update provides an exact solution to the defined editing objective. It should not be interpreted as a global guarantee over all unseen samples or clinical deployment scenarios. We will revise the manuscript to make this scope explicit and tone down absolute claims.

## Experiments

R3: Requested causal tracing ablations and domain shift evaluations.

R: We thank the reviewers for these constructive suggestions. We agree that causal tracing ablations and explicit domain-shift evaluations would further strengthen the empirical analysis. In this work, we focused the limited space on validating the core editing objective across multiple medical imaging benchmarks, emphasizing whether X-Edit can correct targeted errors while preserving test-set performance. Due to the page limit, we did not include exhaustive ablations on the causal tracing module or broader domain-shift protocols. We will clarify this limitation in the revised manuscript and consider these evaluations as important future work.

R2 & R3: Suggested comparing against additional editing and continual learning baselines.

R: We appreciate the reviewers’ suggestion. In the current submission, we compared X-Edit with representative baselines including fine-tuning, fine-tuning with L2 regularization, EWC, LWE, and retraining, covering standard adaptation, continual-learning-based preservation, ViT editing, and an empirical upper bound. Due to the limited space, we were unable to include a more exhaustive set of editing and continual learning methods. We will acknowledge this limitation more clearly and extend the comparison to broader baselines in future work.




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    All the three reviewers agree that the paper deals with an important and clinically relevant problem and proposes a technically sound solution grounded in theory. They have also suggested a few useful suggestions for improvement.



back to top