acceptodds
Under review as a conference paper at ICLR 2027

Competitive Repair Capacity: When Valid Neural Edits Stop Competing

Abstract

Model editing certifies that repair is possible with geometric quantities: editable rank, null- space dimension, codebook size. It leaves a second question open: an acceptable, successful edit can exist and still lose to an unacceptable alternative under the objective that selects edits. We define competitive repair capacity, the number qof valid successful actions beating every unacceptable one, and characterise it exactly: in finite families qis the deletion distance to optimizer-facing failure, its margin-refined form sharply characterises robustness to simultaneous action loss and bounded score error, and continuous families obey a volume law in which admissible dimension alone cannot fix capacity. Two real systems show these are not redundant. On a published GPT- 2 XL AlphaEdit locus competitive capacity collapses to zero across a one-dimensional increase in null-space capacity; one GRACE reproduction is aligned at its released radius, fragile under an off-default geometry stress. An equation- and code-level audit of 41 protocols finds none exposing competitive capacity. A theorem-derived certificate then isolates aligned decisions prospectively — zero observed competitive failures among 141 certified states — yet abstains too often to deploy. Together these results separate the existence of repair capacity from its competitive, certifiable and deployable forms.

Primary area:
foundation or frontier models, including LLMs
Keywords:
model editing, knowledge editing, competitive repair capacity, robustness, large language models

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_gzJxOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Unfortunately, I could barely grasp what this paper is about. It works on model editing. The problem it adresses is a mismatch between the objective model editing optimises and the actual goal of model editing. I did not understand why this issue arises and what this paper does to address the issue.

Critical strengths

Regrettably, I can not speak about strengths, since I did not properly understand what the paper does or is about.

Critical weaknesses

The presentation is unacceptable. The paper is laden with unintroduced jargon. The text does not appear to have a logical structure.

Questions for authors
  • In plain language, what is the issue you adress?
  • In plain language, why is this important?
  • In plain language, what are your contributions?
  • How can a count be negative? (first page, final paragraph)
Reviewer_6ZexOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies a gap between the existence of an acceptable model edit and its selection by an optimization objective. It defines competitive repair capacity as the number of valid successful actions that score above every unacceptable alternative. It characterizes this quantity under action deletion and score perturbation, extends it to continuous families, and examines its behavior in AlphaEdit and GRACE, including a geometry-based deployment certificate.

Critical strengths

The paper gives a precise framework for distinguishing available repairs from repairs favored by the selection objective. Its finite-family characterizations connect this distinction to robustness, and the two editing case studies examine different mechanisms. I appreciate the explicit reporting that the proposed certification did not produce an improvement in deployment utility.

Critical weaknesses

The main observation is a familiar limitation of proxy optimization, as the paper itself acknowledges. The deletion-radius result largely follows from counting the good actions ranked above the best bad action, and the bounded-score-error extension uses a standard two-sided perturbation argument. I do not find that these results, or their organization under a new capacity measure, provide sufficient additional insight.

The abstract formulation places few restrictions on the score. If validity is observable during selection, it can be imposed as a constraint. When it is unavailable, the relevant issue is whether the proposed framework provides useful information from available observables. The reported certificate does not demonstrate a deployment benefit, leaving the practical value insufficiently established.

There is also a local mathematical error in the dimensional comparison following Theorem 4. The claimed volume multiplier omits the ratio of unit-ball volumes, so the stated condition for volume to decrease with dimension does not follow. This requires correction in both the theorem discussion and Appendix A.

Presentation and submission issues include the nonstandard template, the missing required AI use statement, and the extremely small panels in Figure 1. Even setting these aside, I find the contribution insufficient for acceptance. My recommendation is 1 — Reject.

Questions for authors

What additional insight or practical decision capability does competitive repair capacity provide beyond identifying a mismatch between the selection score and validity? In particular, how can it inform selection when validity labels are unavailable, given the reported lack of deployment improvement?

Could the authors correct Theorem 4’s dimensional comparison? For H=hIH=hI, the ratio between successive volumes includes Vol⁡(Bd+1)/Vol⁡(Bd)\operatorname{Vol}(B_{d+1})/\operatorname{Vol}(B_d). For example, h=1h=1 and Δε=1/4\Delta_\varepsilon=1/4 satisfy the claimed shrinkage condition, yet V1=2<V2=π/2V_1=\sqrt{2}<V_2=\pi/2.

Reviewer_EEtJOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

I did not understand anything from this paper. This feels like an AI-generated work report, similar to the likes generated by recent AI agentic models.

Only information seemed clear to me was that the work tried to show despite having representational or parametric capacity for edits (e.g., through null space on Hessian/Activation covariance), an edit may fail. That being said, I believe the paper is trying to claim more beyond the point I made in the last sentence, none of which seemed intelligible or useful to me in how it was presented.

Critical strengths

The question the work asked seems interesting.

Critical weaknesses

Presentation is the first one (random audits, keywords, works without citation (e.g., CRISPEDIT, DSCA, MEMOIR). After that is fixed, I am happy to elaborate.

Questions for authors

Can you write the paper in a more human-readable form?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.