Learned model editors are small networks trained to turn the gradient of one fact correction into a weight change for a large language model. MEND is the standard example. They are trained with a two-level (bilevel) objective. The inner step applies the predicted edit. The outer step checks that the edit worked, and that the model's predictions on unrelated text did not move. That second check, a KL divergence (a measure of how far two output distributions differ) on extra "keep-unchanged" text, is expensive. Step More (EMSEdit) reports it takes over 40% of training time, and replaces it with a plain penalty on the size of the weight change. That penalty is cheap, but it treats every weight direction as equally dangerous. The continual-learning literature, starting with Elastic Weight Consolidation, shows this assumption is wrong.
This note proposes a middle path. To second order, the KL term the editors pay for is a quadratic form in the weight change, weighted by the Fisher information: how sharply the model's outputs react to each weight direction. For a single linear layer, that Fisher factors into two small matrices. One is the covariance of the layer's inputs on preserved text. The other is the covariance of output-side gradients. Both can be computed once, before training. The locality loss then becomes a cheap matrix expression with no extra forward passes. The same factors also give, in closed form, the worst-case input direction for leakage, so we get an adversarial version of locality without an attack loop.
Core Claim
Let $\Delta W$ be the predicted update to an edited layer. Let $A=\mathbb{E}[kk^\top]$ be the second moment of that layer's inputs ("keys") on preserved text. Let $G=\mathbb{E}[gg^\top]$ be the second moment of the gradient $g$ of the log-likelihood with respect to the layer's output, with labels sampled from the model rather than taken from data. Under the standard Kronecker factorization of the Fisher, the locality KL satisfies
$$\mathrm{KL}(p_{W_0}\,\|\,p_{W_0+\Delta W}) \approx \tfrac12\,\big\|G^{1/2}\,\Delta W\,A^{1/2}\big\|_F^2 .$$
We claim three things. First, training a meta-editor with this average-case surrogate in place of the auxiliary KL recovers KL-level specificity at roughly the wall-clock cost of EMSEdit's plain size penalty. Second, the worst-case version, $\|G^{1/2}\Delta W A^{1/2}\|_2^2$, is a meaningful and cheap proxy for adversarial leakage once the accumulated update has rank above one. This spectral norm is the largest whitened output change over unit-whitened key directions, so it is the loss under the most damaging preserved input direction, found by a few power iterations instead of an iterative attack. Third, the two existing penalties are special cases that each drop a factor. EMSEdit's penalty sets $A=G=I$. AlphaEdit's constraint, which confines the update to directions preserved keys never use (their "null space"), keeps only $A$ and enforces it as a hard rule. The output-side factor $G$ is what both miss.
Two caveats limit the claim. The approximation is second order, so it may fail for large edits. The factors are measured at the unedited weights, so they may go stale over long sequences of edits. Both are tested directly below.
Research Program
The claim rests on three strands of prior work, each playing a different role.
MEND and EMSEdit supply the setting. MEND trains a hypernetwork, through meta-gradients, to map rank-one fact gradients to edits, using an outer KL on auxiliary text. EMSEdit removes that KL for speed and adds multi-step backpropagation, and it observes a loss of specificity as inner steps increase. We keep their architectures and change only the locality term, so any difference is attributable to that term.
Three papers motivate using curvature rather than size, and also constrain how it is computed. EWC protects weights that have high Fisher information. HSE brings a diagonal Fisher penalty into sequential editing. It is not a learned editor, and it drops cross-parameter terms. On the Computation of the Fisher Information in Continual Learning warns that cheap approximations hurt. On Split CIFAR-10, the empirical Fisher (built from true labels) reaches about 83.28% and the batched Fisher about 83.38%, against about 84.9% for the exact Fisher. The batched variant also needs very different regularization strengths. For that reason we build $G$ from labels sampled from the model, and we report sensitivity to the regularization weight.
AlphaEdit supplies the input-side factor. It already estimates $A$ from Wikipedia text and projects edits into its null space. Our contribution relative to AlphaEdit is to add $G$, to use the combined factor as a soft differentiable training loss for a learned editor, and to add a worst-case variant. PRUNE links sequential collapse to exploding singular values of the edited weights and clips them heuristically. Our spectral penalty is a principled cousin of that clipping, measured in Fisher-whitened coordinates, and we will test whether it also controls the condition number.
The build has three parts. First, compute $A$ and $G$ once per edited layer on a fixed preserved corpus. Second, add the surrogate to the outer loss of MEND-style and EMSEdit-style editors, trained on zsRE and CounterFact with GPT-J and LLaMA-3 8B. Third, for sequential editing, apply the spectral penalty to the accumulated update, and refresh $A$ and $G$ every $T$ edits. Refreshing tests staleness, and $T=\infty$ is one arm of that test.
Baselines
The naive reference is the unregularized editor: the same hypernetwork trained on the edit loss alone. It shows how much any locality term buys.
The cheapest serious competitor is EMSEdit's size penalty, run with its published architecture and the same number of inner steps. Beating it means clearly higher specificity (neighborhood accuracy on CounterFact and zsRE) at efficacy within about a point, at comparable training time. EMSEdit reports 95.40% efficacy and 93.91% generalization for sequential zsRE editing on LLaMA-3 8B. It also reports an average score of 72.07 on 8,192-edit CounterFact batches, against 53.04 for MALMEN. I found no reported specificity number for EMSEdit on Lacuna, so we will measure it ourselves.
The strongest quality competitor is MEND's original auxiliary-KL training, given the same data and a wall-clock-matched run. Our method wins if it reaches equal specificity with at least 30% less training time, which is plausible given the reported 40% KL share.
For sequential editing, the external references are AlphaEdit, which reports sustained efficacy and fluency over 3,000 sequential updates, and HSE, which reports GLUE retained after 1,000 edits where MEMIT and MEND fall to near zero. We run both on the same edit streams.
Three ablations isolate the key ingredient. The first uses only $A$ as a soft penalty. The second uses only $A$ as a hard AlphaEdit-style projection of the hypernetwork's output. The third uses a diagonal Fisher. These separate the value of $G$ from the value of the Kronecker structure. Every arm receives the same regularization sweep of eight values, chosen on a held-out validation split.
Why not just…?
Why not just project the editor's output into AlphaEdit's null space? This is the most serious threat. It is free, it needs no training change, and it provably leaves preserved keys untouched. The weakness is that the null space of $A$ can be small for the busy middle layers where editing matters. In that case the projection leaves little room for efficacy. It also cannot tell which remaining directions matter for the outputs, which is exactly the information $G$ carries. The hard-projection ablation settles this. If it matches our penalty on specificity and efficacy across CounterFact, zsRE and long sequences, the output-side factor adds nothing, and we will report that.
Why not just use a diagonal Fisher, as EWC and HSE do? The diagonal is cheaper still. However, an edit to a weight matrix is structured as an outer product, and the diagonal throws away the input and output correlations that give that structure its meaning. The Fisher-computation study shows that approximation quality measurably changes retention. The diagonal-Fisher ablation, run at equal tuning budget, tests whether that finding carries over to editors.
Why not just shrink the auxiliary KL batch? Evaluating the KL on fewer tokens could recover most of the speed at little cost. We include a KL arm with its auxiliary batch reduced until its wall-clock time matches our method. If that arm matches us, the surrogate's advantage is only convenience.
Why not attack edited models with discrete token search, or use sharpness-aware training as CoRSA does? CoRSA gains 12.42% in generalization and cuts forgetting by 27.82% relative to LoRA. It does so by adding an extra gradient-ascent step for every update, which doubles the cost of each step. Its adversary also perturbs weights, not preserved inputs. Token-level attacks are the ground truth for leakage, but work on NLP adversarial training documents their very high query cost. We therefore use such attacks only for evaluation, and ask whether our closed-form spectral score predicts them.
Proposed Experiments
The first decisive check tests the surrogate directly. On held-out edits, we correlate the surrogate with the true KL on preserved text, binned by edit size. This tells us where the second-order approximation stops tracking the true KL.
The second check is the specificity-versus-training-time comparison of all baseline arms on CounterFact and zsRE, with GLUE, GSM8K and Wikitext perplexity as collapse monitors.
The third check runs sequential streams of 100, 1,000 and 3,000 edits. For each, we track specificity, GLUE, the condition number of the edited matrices (as studied by PRUNE), and the benefit of refreshing $A$ and $G$ against a single computation at the start.
The fourth check tests the adversarial proxy, using the distract-neighborhood evaluation of SADR plus a budgeted token-substitution attack that searches for prompts that leak the edited fact. We ask whether edits with a lower spectral score leak less under attack. We also compare the spectral penalty against the trace penalty and against SADR's attention regularizer.
If the spectral score does not predict attack success, the adversarial claim fails, and the proposal falls back to its first claim alone.
References
- Step More: Going Beyond Single Backpropagation in Meta Learning Based Model Editing
- Overcoming catastrophic forgetting in neural networks
- Fast Model Editing at Scale
- Hippocampal-like Sequential Editing for Continual Knowledge Updates in Large Language Models
- On the Computation of the Fisher Information in Continual Learning
- AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models
- Perturbation-Restrained Sequential Model Editing
- Conflict-Resolving and Sharpness-Aware Minimization for Generalized Knowledge Editing with Multiple Updates
- Towards Improving Adversarial Training of NLP Models
- Revealing and Mitigating Over-Attention in Knowledge Editing