- Rating
- 1
- Confidence
- 2
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 2(+1)
Read review
Paper summary
This paper essentially works on knowledge editing for LLMs, in the locate-then-edit setting of ROME/MEMIT/AlphaEdit. An edit is a weight change D to a single linear layer (an MLP output projection). It should (i) move the layer's outputs at the edited keys toward desired values (ii) barely change the outputs on a set of "protected" keys and (iii) keep the update small. The question the paper asks is how to decide whether such an edit exists when the covariance matrix Q is not stored exactly. Instead, Q is known only through a Frequent Directions sketch with a residual bound, plus a few exact covariance-vector products Qv.
The paper makes three theoretical points. First, two streams of protected keys can produce identical Frequent Directions sketches but opposite answers about whether a given edit request is feasible (Proposition 1), so a sketch alone cannot always decide. Second, given exact products Qv on a set of directions, classical PSD completion yields the tightest lower and upper bounds on the unknown part of Q, uniformly over all candidate edits (Proposition 2). Third, combining an edit that is feasible under the upper bound with a weak-duality lower value computed under the lower bound gives a three-way decision (Proposition 3): a "witness" (the request is feasible), an "exclusion" (no edit satisfies all three budgets), or "unresolved". The authors also discuss batch edits (one update serving several targets) and sequential edits, where the budgets must be checked on the cumulative update.
The GPT-2 experiments are illustrative. They show that all three outcomes can occur on a real editing problem, and that "unresolved" differs from "infeasible". They also show that covariance-vector queries are needed beyond the sketch, but the proposed query policies perform about the same as a Krylov baseline. Finally, a witness for the layer-level budgets does not guarantee preserved model behavior on unrelated facts.
Critical strengths
- source code is provided which is good for reproduction.
- simple and correct technical results. The bounds are easy to verify.
Critical weaknesses
- Very poor readability. The paper is hard to follow. To name a few,
-
The introduction never states the concrete problem. It speaks of "modifying a model at a linear projection", "protected feature rows" and "requested responses", but never says that the edit is a weight update to a single MLP output projection of a transformer, what the protected features and new features are (MLP inputs at calibration positions and at the edited subject tokens), or where the requested responses come from (target output vectors optimized as in ROME). The word "MLP" first appears in Section 3, and the layer and model actually edited only in Section 5 and Appendix B. Nor does the introduction explain in what practical situation the key covariance cannot be stored, which is the premise of the entire paper. The reader has to reverse-engineer from Sections 2-5 and the appendix that the setting is locate-then-edit knowledge editing as in ROME/MEMIT.
-
Many technical terms in the main text are used without definition. Examples:
- "accumulated shrink values" and "residual traces" (Introduction, Proposition 1). These are Frequent Directions internals, and Proposition 1 depends on them, but the FD update is never described.
- "complete protected prefixes, including answer paths" and "untied projection" (Introduction).
- "upper-feasible candidate", "full-residual-corrected lower value" and "witness", all used in the Introduction before Section 2.4 defines them.
- "native" evaluation versus non-native, "physical response units" (Sections 4-5, Figure 1).
-
The setting is standard locate-then-edit knowledge editing, but the paper renames nearly every standard concept without relating it to the established terms. The "protected covariance" Q is the uncentered key covariance used by ROME () and MEMIT (), and the "protected response cost" is their term for preserving previously stored associations. The new features K and requested changes R coincide with MEMIT's new keys and residual . "Shared responses" is batch (mass) editing as in MEMIT, "cumulative admission" is sequential editing as studied by AlphaEdit, and "rewrite/paraphrase/neighborhood preference" are CounterFact's efficacy, paraphrase and neighborhood scores (ES/PS/NS in ROME). The term "protected features" also suggests a privacy or fairness setting, which is not what the paper is about.
- The setting is not realistic: storing Q is feasible, so the motivation for sketching is unclear.
The paper motivates the sketch by the quadratic cost of storing the key covariance . At the scale studied, this cost is small:
-
Q for the edited GPT-2 layer is 3072 x 3072, about 38 MB in fp32. ROME/MEMIT routinely precompute and store exactly this matrix. Even for MLP widths of order 10^4 in larger models, the covariance of one layer remains well within a single accelerator's memory.
-
In the experiments, Q is built from only n = 512 protected positions. Storing A itself (512 x 3072, which the paper reports as 12.6 MB) gives Q exactly. More generally, exact Q is always available at a storage cost of . The query policies store 4.6-7.2 MB, so the memory saving over exact information is only about 2-3x.
-
The method is not actually sketch-only. Each covariance-vector action requires another full pass over the protected features, and the experiments use up to 33 passes. So the method still needs either A or the ability to recompute A by rerunning the model. The implied regime (Q too large to store, yet repeated passes over the protected features affordable) is never motivated or measured with a concrete memory or compute budget.
-
The paper does not compare against the obvious alternative of simply storing Q, or A, and solving the editing problem exactly, as ROME/MEMIT/AlphaEdit do. Without such a comparison, or experiments in a regime where Q genuinely cannot be stored (larger models, many protected keys, a tight memory budget), it is unclear when the proposed interface would be preferable in practice.
Questions for authors
In what concrete setting is storing the key covariance Q infeasible while a sketch plus covariance-vector queries is affordable? For the GPT-2 layer studied, Q is 3072 x 3072 (about 38 MB in fp32), and it is built from only 512 protected positions, so storing A itself (12.6 MB as reported) gives Q exactly. Could you specify the model size, number of protected keys, and memory budget where the proposed interface is preferable, and report measurements in that regime?
2 comments
Thanks for your response. I have increased the soundness score, but my concern over real scenario improvement is not addressed, and I would like to keep my evaluation.
Reviewer tRrf
The full-information comparison is important for understanding what is lost when covariance information is incomplete. The submitted paper includes objective references for this comparison. Section 5.1 reports the full-information target errors for the shared-response problem, and Figure 1 plots them against the decision intervals. Section 5.2 uses 96 full-information scalar solutions as objective references across the 32 factual requests and three protection caps. At relative tolerance 0.5 in Figure 1, the squared threshold is 1235.766 and the intermediate request has a full-information target error of 1200.896. Its retained-information interval is [1175.352, 1963.695], so the interface correctly leaves it unresolved.
To make the setting behind these comparisons concrete, the editing operation changes the weight matrix of one transformer MLP output projection. Protected keys are MLP inputs at calibration positions, new keys are inputs at the editing positions, and target responses specify the desired output changes. For the factual study, these targets are optimized latent vectors following the locate-then-edit construction. Equation (1) asks whether one update meets the target-fit, calibration-response and update-norm budgets jointly. We will put this example and its connection to standard key-covariance notation in the introduction, and define the FD accumulated shrink and residual trace when first used.
The information constraint has a precise consequence even when a full-information solve is available for evaluation. Proposition 1 constructs two realizable streams with identical FD sketches, accumulated shrink and residual trace, for which the same request has opposite feasibility answers. Proposition 2 then characterizes the protection costs consistent with the retained radius and acquired columns, uniformly over all candidate updates. Together these results establish what additional covariance information can settle and why a sound decision rule sometimes returns unresolved.
The cost of accessing this information is central to the practical question. For the measured GPT-2 layer, retaining the 512 × 3072 FP64 feature array costs about 12.6 MB, the dense FP32 covariance costs about 37.7 MB, and the policy uses up to 33 feature passes including FD setup. The intended operating regime has a bounded resident covariance state and access to streamed or recomputed features. The submitted objective references quantify the effect of incomplete information on the editing solution.
Finally, the review describes the technical results as “simple and correct” and the bounds as easy to verify, while assigning soundness 1. Could the reviewer identify the specific mathematical statement or proof step underlying that rating? This would let us address the validity concern directly.