acceptodds
Under review as a conference paper at ICLR 2027

HoReN: Normalized Hopfield Retrieval for Large-Scale Sequential Model Editing

Abstract

Large language models encode vast factual knowledge that can become outdated or incorrect after deployment, yet retraining is prohibitively costly. This motivates lifelong model editing, which updates targeted behavior while preserving the rest of the model. Existing editors, both parameter-modifying and parameter-preserving, degrade severely as edits accumulate and struggle to generalize across paraphrases. We propose HoReN, a codebook-based parameter-preserving editor that wraps a single MLP layer with a discrete key–value memory. HoReN treats each codebook entry as both a knowledge key and a Hopfield stored pattern, retrieves edits by angular similarity on the unit hypersphere, and refines queries through damped Hopfield dynamics so paraphrases move toward the correct memory basin while unrelated inputs remain stable. HoReN achieves strong editing performance with consistent gains across diverse benchmarks spanning standard ZsRE, structured WikiBigEdit, and unstructured UnKE evaluations. Moreover, HoReN scales to 50K sequential edits on ZsRE with stable overall performance above 0.93, while the evaluated baselines collapse or degrade severely before reaching 10K. Our code is available at https://anonymous.4open.science/r/elephant-4B7F.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Hopfield, lifelong learning, large language model editing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qDJnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes HoReN, a codebook-based parameter-preserving model editing method that models knowledge keys as Hopfield stored patterns and combines angular retrieval on the unit hypersphere with damped Hopfield dynamics to achieve more robust edit retrieval and paraphrase generalization. Extensive experiments are conducted to validate the effectiveness of the proposed method.

Critical strengths

The paper projects keys and queries onto the unit hypersphere, replacing magnitude-sensitive Euclidean matching with cosine similarity.

The experiments in this paper are detailed and thorough.

The written expression in the paper is relatively clear.

Critical weaknesses

Figure 1 shows that, as the number of sequential edits increases, different methods exhibit different forms of overall performance degradation. However, this evidence mainly reflects performance-level degradation and may not be sufficient on its own to attribute the issue specifically to routing.

The treatment of the query over the memory keys in the Hopfield refinement is reasonable; however, I believe there is still room for discussion as to whether the methodological novelty is sufficient to support the relatively strong theoretical framing adopted in the paper.

Since M=1M=1, please compare against generic normalized attention/kernel-smoothing retrieval to clarify the algorithmic novelty.

The theory analyzes the standard unnormalized Hopfield operator, whereas HoReN actually uses a normalized, damped, one-step operator.

Questions for authors

Refer to the above text.

3 comments

Point 1 (Potential Exp) Evidence attributing degradation to routing

Response

We agree that Figure 1 shows how end-to-end performance changes with edit count. To test the role of routing directly, we varied GRACE's initial Euclidean radius on the same 1,000-edit ZsRE stream with LLaMA-3.1-8B-Instruct. We held the backbone, editing layer, pooling, and value-training recipe fixed and rebuilt memory at each setting. Increasing the initial radius from 1.0 to the calibration-selected 6.8173 raised Gen from 6.66% to 33.57%, with Loc at 100% in both runs and Rel above 98%. In the descriptive sweep, GRACE with normalized keys and queries and adaptive cosine-distance radii reached 61.30% Gen at 99.90% Loc.

These results show that routing choices materially affect paraphrase performance in the tested GRACE setup, where original edit prompts already have high reliability. The routing rule can also change which entries are allocated or reused during editing.

Reported routing-calibration results.

Method / settingGen (%)Loc (%)
GRACE, initial Euclidean radius 1.06.66100.00
GRACE, calibrated initial Euclidean radius 6.817333.57100.00
GRACE + Norm + adaptive cosine-distance radii, descriptive sweep61.3099.90

Point 2 (Exp) Novelty over normalized attention and smoothing

Response

Thank you for asking us to clarify the relationship to normalized attention and kernel smoothing. We address the shared mathematical operation, the editing-specific design, and the corresponding controls separately.

1. Relationship to attention and kernel smoothing.

The weighted read u(q)=normalize⁡(∑ipi(q)ki)u(q)=\operatorname{normalize}(\sum_i p_i(q)k_i), with pi(q)∝exp⁡(20q⊤ki)p_i(q)\propto\exp(20q^\top k_i), is a normalized attention read. For unit queries and keys,

exp⁡(20q⊤ki)=e20exp⁡ ⁣(−10∥q−ki∥2),\exp(20q^\top k_i)=e^{20}\exp\!\left(-10\|q-k_i\|^2\right),

so the same weights correspond to a Gaussian kernel with squared bandwidth 1/201/20. The weighted-read operation is shared with these methods. HoReN uses this read within a sequential editor: it blends the direction with the original query using γ=0.1\gamma=0.1, normalizes the result, applies a hard acceptance threshold, and uses the resulting match for both inference and online memory allocation. A normalized-attention implementation with the same blend would implement the same one-step refinement formula. We will frame the contribution around this editing design and its empirical behavior, rather than presenting the shared attention operation as a distinct new primitive.

2. Direct normalized-attention control.

We compare against direct replacement of the routing query by u(q)u(q) on the same 1,000-edit ZsRE stream with LLaMA-3.1-8B-Instruct. Each method builds its own memory. The following are the original reported results, in percentages.

MethodThreshold ccGenLocOP
HoReN0.8592.0499.4596.81
Direct normalized attention0.8513.617.4718.64
Direct normalized attention (fallback calibration)0.91666710.1690.0532.27

The second attention threshold was selected under the fallback calibration target of 75.12% Loc. At the reported test settings, the damped blend achieves higher Gen and Loc than direct replacement. This comparison tests how the weighted read is combined with the original query, including its effect on memory construction. Its result supports controlling the strength of the update; it does not distinguish HoReN from attention methods implementing the same residual blend.

3. Evidence for the memory-construction effect.

Refinement also operates before the editor decides whether to allocate a new key or reuse a same-label entry. We investigated this effect using the same-step Top-1 control discussed in our additional experiments. On a selected 100-example sequence with fixed edit order, five additional initializations give:

MethodRel (%)Gen (%)Loc (%)OP (%)Learned keys, excluding sentinel
HoReN99.9595.2610098.3899
Top-1 interpolation98.9589.2610095.9588

Each run has seven HoReN wins, one Top-1 win, and 92 ties. For those seven winning paraphrases, both routers succeed when given HoReN's memory, and both fail when given Top-1's memory. This swap identifies a memory-coverage effect: the refinement rule changes which keys are stored, with HoReN retaining more entries in this case. These are repeated initializations of one selected sequence, not independent editing streams. The case study uses float32 threshold comparison, while the default-stream controls use the legacy BF16 gate; one of the seven allocation differences is precision-sensitive.

Together, these experiments show how the query update affects thresholded routing and memory construction. We will separate this editor-level contribution from the underlying attention formula and align the framing with the analysis in Point 3.

Point 3 (Theory) Theory for the deployed operator

Response

Thank you for noting the operator difference. As stated in Section 2.3 and Appendices A and D, the original theory motivates limiting memory attraction; it was not presented as a retrieval or locality guarantee for the deployed one-step rule. We have extended the analysis to the actual normalized, damped operator to make the connection explicit.

1. Analysis of the deployed operator.

For a fixed finite codebook, define

\quad T(q)=\nabla F(q), \quad E(q)=\tfrac12\|q\|^2-F(q),$$ and $$u(q)=\frac{T(q)}{\|T(q)\|}, \qquad H_\gamma(q)=\operatorname{normalize}\bigl((1-\gamma)q+\gamma u(q)\bigr).$$ For unit $q$, $T(q)\ne0$, and $0<\gamma<1/2$, the deployed update satisfies $$E(H_\gamma(q))-E(q) \le-\frac{1-\gamma}{2\gamma}\|T(q)\|\, \|H_\gamma(q)-q\|^2\le0.$$ With fixed $\|k_i\|\le1$ and $T(q_0)\ne0$, its ideal repeated iteration also converges to a codebook-induced fixed point. This concerns exact arithmetic before fixed-step or epsilon stopping. The following proof applies directly to $H_\gamma$. **Proof of the added result and displacement bound.** Assume a fixed finite codebook, $\|k_i\|\le1$, $\beta>0$, unit $q_0$ with $T(q_0)\ne0$, and constant $0<\gamma<1/2$. We use exact updates without epsilon stopping. **Energy descent.** Write $g=T(q)=\nabla F(q)$, $u=g/\|g\|$, $v=(1-\gamma)q+\gamma u$, and $q^+=v/\|v\|$. Since $\|v\|\ge1-2\gamma>0$ and $q^+$ maximizes $v^\top z$ over unit $z$, $$v^\top(q^+-q)\ge0 \quad\Longrightarrow\quad \gamma u^\top(q^+-q)\ge\frac{1-\gamma}{2}\|q^+-q\|^2.$$ Convexity of $F$ then gives $$F(q^+)-F(q)\ge g^\top(q^+-q) \ge\frac{1-\gamma}{2\gamma}\|g\|\|q^+-q\|^2.$$ Both queries have unit norm, so their quadratic energy terms cancel, yielding the stated descent inequality. **Uniform nondegeneracy.** The set $S=\{q:\|q\|=1,F(q)\ge F(q_0)\}$ is compact and contains the trajectory. It contains no zero of $\nabla F$: such a zero would be a global minimizer of the convex function $F$, forcing $q_0$ also to be a minimizer and contradicting $T(q_0)\ne0$. Therefore $m=\min_{q\in S}\|T(q)\|>0$, and $$E(q_s)-E(q_{s+1})\ge a\|d_s\|^2, \qquad a=\frac{(1-\gamma)m}{2\gamma},\quad d_s=q_{s+1}-q_s.$$ Thus energies converge and $d_s\to0$. The constant $m$ is not asserted to be uniform across codebooks or edit scales. **Single-limit convergence.** Let $P_q=I-qq^\top$ and $b_s=\|(1-\gamma)q_s+\gamma u(q_s)\|\le1$. Projecting the update onto the tangent space gives $$\|\operatorname{grad}_{\mathbb S}E(q_s)\| =\|P_{q_s}T(q_s)\| \le\frac{\|T(q_s)\|b_s}{\gamma}\|d_s\| \le\frac1\gamma\|d_s\|.$$ The restriction of $E$ to the sphere is real analytic. Equivalently, $\widetilde E(x)=1/2-F(x/\|x\|)$ is an analytic extension near the sphere, whose gradient on the sphere is the above tangent gradient. Its KL inequality, uniformized near the compact accumulation set at limiting energy $E_\infty$, gives a concave desingularizing function $\varphi$. Combining the preceding two inequalities yields, for sufficiently large nonstationary $s$, $$\varphi(E(q_s)-E_\infty)-\varphi(E(q_{s+1})-E_\infty) \ge a\gamma\|d_s\|.$$ Summation proves finite trajectory length, hence convergence to $q_*$. Continuity of $H_\gamma$ on $S$ and $d_s\to0$ give $H_\gamma(q_*)=q_*$. If the energy becomes constant, sufficient descent makes the sequence stationary. This follows the Appendix D KL argument; see [Absil et al.](https://doi.org/10.1137/040605266). **Single-step displacement.** For fixed unit $q$, the vector $(1-\gamma)q+\gamma u$ lies on the sphere with center $(1-\gamma)q$ and radius $\gamma$. Its maximum angle from $q$ is $\arcsin(\gamma/(1-\gamma))$. Consequently, $$\|H_\gamma(q)-q\| \le\sqrt{2-2\sqrt{1-(\gamma/(1-\gamma))^2}},$$ which is approximately 0.111283512 at $\gamma=0.1$. These are fixed-codebook, exact-arithmetic statements; they do not establish task accuracy, finite-precision convergence, or convergence across changing codebooks. **2. Why one step is selected.** Editing also requires rejecting unrelated queries, which the energy does not distinguish from edits. Damping bounds adjustment (0.111284 at $\gamma=0.1$); the step count is selected empirically. Table 12 and Figure 10 report the step-count ablation at 1,000 ZsRE edits on LLaMA-3.1-8B, with $\beta=20$ and $\gamma=0.1$: | Refinement steps M | Gen (%) | Loc (%) | Reported OP (%) | |---:|---:|---:|---:| | 1 | 92.6 | 99.3 | 97.0 | | 2 | 92.8 | 94.3 | 95.5 | | 4 | 93.0 | 52.1 | 78.4 | One step has the highest reported OP among these tested step counts. Increasing from one to four steps improves Gen by 0.4 percentage points while reducing Loc by 47.2 points. Thus, the theory motivates limited attraction, damping controls its magnitude, and the ablation selects the one-step operating point. We will make this separation explicit in Section 2.3 and include the deployed-operator analysis and proof.

Thank you for your detailed point-by-point responses. After thoroughly consideration, I have decided to maintain my original rating.

Reviewer_6fEnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes HoReN, a parameter-preserving sequential model editing method. It follows the structure of GRACE: freezing the base model and attaching a key-value codebook on a single-layer MLP. The authors make two changes to the routing rules: first, they project queries and keys onto a unit hypersphere and use a cosine threshold to determine hits; second, before matching, they apply a damped Hopfield update to the query (β=20, γ=0.1, M=1). The keys written into the codebook are fixed as the original unrefined queries. The theoretical part proves the energy descent and convergence properties of standard Hopfield iterations, used to argue against multi-step iteration. Experiments cover 7 models and 3 benchmarks: on ZsRE with 1K edits, OP is 0.94–0.97; on LLaMA-3.1-8B with up to 50K edits, OP remains around 0.93, while all baselines either collapse or degrade significantly within 10K edits.

Critical strengths
  1. Strong and stable results under large-scale sequential editing. HoReN maintains high Rel/Gen/Loc simultaneously across multiple model families and edit scales; on LLaMA-3.1-8B with 50K edits, all three metrics remain above 0.89, whereas baselines exhibit cliff-like drops, persistent drift, or near-zero generalization. GRACE, WISE, and HoReN are aligned to the same editing layer, ruling out confounds from layer selection. Sequential editing collapse beyond 10K edits is a core challenge in this field; if these results are reproducible, they hold substantial practical value.

  2. Minimal method with clear diagnostics. The authors localize the failure of GRACE-style methods to routing selectivity, and the step-by-step ablations and locality sensitivity analyses (Table 12/14/15) jointly characterize the generalization–locality trade-off. Hyperparameters are fixed across models, and the overhead is comparable to GRACE, facilitating reproduction and deployment.

Critical weaknesses
  1. The Hopfield step may be largely equivalent to lowering the threshold, lacking discriminative controls. When the softmax concentrates on the correct key, the refined cosine is a strictly monotonic function of the original cosine; therefore, HoReN's hit decision under γ=0.1, c=0.85 is equivalent to GRACE+Norm's decision under a threshold of approximately 0.818. Table 12 shows that at γ=0.05 (equivalent threshold ≈0.835), Gen remains 0.911, while GRACE+Norm at the same threshold is only 0.33. This implies that a large number of rewritten sentences have cosines falling within a narrow band just below the threshold, and GRACE+Norm could approach HoReN with only a slight threshold adjustment. The paper does not provide this threshold sweep, which directly determines the actual magnitude of the contribution.

  2. The evaluation evidence is insufficient to support the core claim of "large-scale stability." The OP across five edit streams is 0.962±0.003, yet the default stream used in the main table is 0.971, more than 3 standard deviations higher; the main-table results may come from a favored edit stream. The 50K experiment appears to include only HoReN with a single edit stream, and it is unclear how the rewritten sentences and locality queries for the additional ZsRE entries were constructed. ZsRE's locality queries are naturally separated from edit queries, and the threshold c and pooling ratio were selected based on ablations on ZsRE; on WikiBigEdit, where locality is harder, Loc is only 0.63–0.68.

Questions for authors
  1. Please provide the Gen–Loc curve of GRACE+Norm under threshold c∈[0.70, 0.95], plotted together with HoReN in the same figure. If HoReN's Gen remains significantly higher at the same Loc, I will raise my score.

  2. Please report the cosine distributions between rewritten sentences and the correct key, and between locality queries and the nearest key, before and after refinement, as well as the softmax weight of the correct key under different N. A larger codebook should dilute this weight, yet Gen remains roughly flat—what is the reason?

  3. Is the default stream within the five edit streams? Why is its result significantly above the five-stream mean?

  4. How were the rewritten sentences and locality queries for the additional entries in the 50K data constructed? Were c and the pooling ratio selected on data non-overlapping with the evaluation set?

The experimental breadth and stability are highlights, but the Hopfield step can be reduced to lowering the threshold when softmax is concentrated, and no control currently rules out this explanation; the representativeness of the main-table and 50K results is also in question. If the rebuttal adds the threshold sweep, explains the discrepancy between the default stream and the five-stream results, and clarifies the construction of the 50K data, I am willing to raise my score to 3

3 comments

Point 1 (Exp) Threshold sweep and matched-locality comparison

Response

Thank you for suggesting this control. We swept HoReN's fixed cosine threshold and the initial radii of two normalized GRACE variants at 1,000 edits on the same LLaMA-3.1-8B-Instruct/ZsRE stream. Both GRACE variants retain GRACE's adaptive per-key radii and its allocation, reuse, and conflict rules. All runs use layer 29, normalized keys and queries, and mean pooling over the final 60% of prompt tokens. Each setting builds a fresh memory. The GRACE runs retain the implementation's Adam optimizer, learning rate 1, and 100 update steps.

The tables cover the requested display range [0.70,0.95][0.70,0.95]. For HoReN, the displayed value is its fixed cosine threshold cc. For Euclidean GRACE, C=1−rC=1-r is a plotting convention that orders settings from looser to tighter initial radii. For cosine GRACE, which uses distance 1−cos⁡(q,k)1-\cos(q,k), C=1−rC=1-r is the initial cosine boundary. In both GRACE variants, the per-key radii change during editing. Dashes indicate settings with no completed run for that method.

GRACE + Norm + Euclidean

Displayed CCHoReN Gen (%)HoReN Loc (%)GRACE initial radiusGRACE Gen (%)GRACE Loc (%)
0.70094.328.450.30028.18100.00
0.75094.3217.590.25021.73100.00
0.80094.7057.680.20015.92100.00
0.818——0.18213.52100.00
0.835——0.16511.82100.00
0.85092.0499.450.15010.42100.00
0.90074.24100.000.1006.16100.00
0.95038.83100.000.0505.06100.00

GRACE + Norm + cosine

Displayed CCHoReN Gen (%)HoReN Loc (%)GRACE initial radiusGRACE Gen (%)GRACE Loc (%)
0.70094.328.450.30061.1099.90
0.75094.3217.590.25061.1099.90
0.80094.7057.680.20061.3099.90
0.818——0.18261.1099.90
0.835——0.16560.9099.90
0.85092.0499.450.15060.6099.90
0.90074.24100.000.10056.62100.00
0.95038.83100.000.05035.79100.00

Matched locality at 1,000 edits

For each method, the table uses the measured point with the highest Gen among the displayed settings that achieved exactly 100% Loc. This is the only measured locality value shared by HoReN and both adaptive GRACE variants in these sweeps.

Matched Loc (%)ComparisonHoReN ccHoReN Gen (%)GRACE initial radiusGRACE Gen (%)HoReN − GRACE Gen (pp)
100.00GRACE + Norm + Euclidean0.9074.240.3028.18+46.06
100.00GRACE + Norm + cosine0.9074.240.1056.62+17.62

Cosine GRACE also reached 61.30% Gen at 99.90% Loc at initial radius r=0.20r=0.20 in the earlier coarse sweep. In the extended Euclidean sweep, r=0.40r=0.40 falls outside the requested display range and reached 33.24% Gen at 100% Loc. The matched-locality results compare the complete editors, whose memory-update rules differ.

Your threshold argument applies when the memory is fixed and the refinement direction is the closest key. We also tested normalized cosine matching with refinement disabled, using HoReN's memory-update and value-training rules. At fixed c=0.818c=0.818 and c=0.835c=0.835, this control reached 93.13% Gen / 96.33% Loc and 92.44% Gen / 99.55% Loc, respectively. HoReN at c=0.85c=0.85 reached 92.04% Gen / 99.45% Loc. At 100% Loc, the no-refinement control reached 88.88% Gen at c=0.85c=0.85, versus HoReN's best observed 74.24% at c=0.90c=0.90. These descriptive results show that the fixed-threshold control can recover HoReN's trade-off on this stream. The gains over adaptive GRACE at 100% Loc therefore describe the performance of the complete editors on these tested settings.

No-refinement control: reported operating points.

MethodFixed threshold cGen (%)Loc (%)
HoReN0.8592.0499.45
No refinement0.8588.88100.00

Point 2 (Exp) Cosine distributions and softmax dilution

Response

In the standalone ZsRE run, we measured cosine scores before and after refinement and designated-key softmax weights at 1K, 10K, and 50K actual edits. The table reports medians; complete distributions are provided in the supporting experiment report. The designated key is the training-time slot and its softmax weight is measured before refinement.

Actual editsParaphrase → designated cosine, before / afterLocality → nearest cosine, before / afterMedian target weight (%)Median target margin
1,0000.9297 / 0.93750.7773 / 0.80086.01810.08594
10,0000.9258 / 0.92970.7891 / 0.81250.55850.05859
50,0000.9219 / 0.92970.7969 / 0.82030.11140.04688

The median designated-key weight fell from 6.0181% to 0.1114% as the codebook grew, while remaining about 56–60 times the uniform weight. HoReN normalizes the weighted key sum before blending it with the original query. This cancels the common softmax denominator in exact arithmetic. A smaller individual weight therefore does not by itself make the refinement smaller. Additional keys can still change the direction of the weighted sum and compete in the final match.

The median paraphrase-to-designated-key cosine increased after refinement at every measured scale. At 50K, 9,031 paraphrases achieved full target-token accuracy after initially routing to a different slot. These observations help explain why Gen remains high despite dilution. The measurements do not isolate how much each effect contributes. From 1K to 50K, Gen fell from 94.17% to 90.52%, and Loc fell from 99.40% to 91.59%.

Tabulated cosine-distribution summaries. Values are taken from the saved statistics for the same diagnostic reproduction. Cosines are on the 0–1 scale; P05 and P95 are the 5th and 95th percentiles. These summarize the plotted distributions rather than histogram-bin densities.

Actual editsQuery/key pairRefinementMeanP05MedianP95
1,000Paraphrase / designated keybefore0.9178520.8085940.9296880.988281
1,000Paraphrase / designated keyafter0.9245270.8242190.9375000.988281
1,000Locality / nearest keybefore0.7695780.7031250.7773440.820312
1,000Locality / nearest keyafter0.7957500.7382810.8007810.839844
10,000Paraphrase / designated keybefore0.9146480.8085940.9257810.988281
10,000Paraphrase / designated keyafter0.9212550.8242190.9296880.988281
10,000Locality / nearest keybefore0.7846500.7187500.7890620.835938
10,000Locality / nearest keyafter0.8100160.7500000.8125000.855469
50,000Paraphrase / designated keybefore0.9140180.8085940.9218750.988281
50,000Paraphrase / designated keyafter0.9205950.8242190.9296880.988281
50,000Locality / nearest keybefore0.7904660.7226560.7968750.839844
50,000Locality / nearest keyafter0.8148180.7578120.8203120.859375

Tabulated softmax-dilution results. K includes the sentinel entry. Weights are percentages; the last column is a dimensionless ratio. These values belong to the same diagnostic reproduction as the distribution measurements above.

Actual editsKMedian designated-key weight (%)Uniform weight 1/K (%)Median / uniform
1,0001,0006.0180660.10000060.1807
10,0009,9800.5584720.01002055.7355
50,00049,8860.1113890.00200555.5676

Point 3 Default stream versus five-stream results

Response

No. The default stream was outside the five-stream aggregate. At 1,000 edits, it reached 97.09% OP, compared with a five-stream mean of 96.17% and a sample standard deviation of 0.30 percentage points. The gap mainly reflects higher Gen on the default stream (92.6% versus 90.6% across the five streams). We will report both results together.

Point 4 (Potential Exp) 50K evaluation validity and data construction

Response

The additional entries come from ZsRE. As described in Appendix E.4.4, the 50K experiment uses a pool of 75,658 records, combining the standard released split with additional ZsRE entries that do not overlap with that split. The edit prompts, target answers, paraphrases and locality queries are all taken from ZsRE, with the associated queries retained for each edit. We will make their source and this correspondence clearer in the experimental description.

Both the matching threshold cc and the token pooling ratio were selected on data disjoint from the evaluation set. They were selected once per model and then held fixed across datasets and edit scales, including the 50K experiment. We will state this separation explicitly in the paper.

Thank you very much to the authors for their detailed, point-by-point responses. However, given my overall concerns regarding the space complexity of the technical line this work belongs to—namely, external storage-based model editing methods—I need to consider further and discuss with other reviewers before I can finalize my score.

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

This paper proposes HoReN, a codebook-based parameter-preserving editor built on an external key–value memory. It builds on a GRACE-style codebook, normalizes queries and memory keys onto the unit hypersphere, and refines the query representation through a single damped Hopfield-style update. A similarity threshold then determines whether the matched value is added to the layer output. The authors evaluate seven models on ZsRE, WikiBigEdit, and UnKE, and report results with up to 50K sequential edits. HoReN maintains high Reliability, Generalization, and Locality at large editing scales, substantially outperforming the baselines evaluated in the paper.

Critical strengths
  1. The method is simple and practically appealing. HoReN leaves the base model’s weights unchanged and modifies only the routing mechanism of the external codebook, making it easy to integrate into existing GRACE-style editors.
  2. The experimental coverage is broad. The evaluation spans seven models, three datasets, multiple model families, and streams of up to 50K edits, exceeding the scale of many prior studies that evaluate only a single model or relatively short editing streams.
  3. The component ablations are relatively comprehensive. Table 14 separates the contributions of normalization and Hopfield refinement, while the appendices further analyze the number of iterations, damping factor, token pooling, value adaptors, and false-positive memory-activation rates.
Critical weaknesses
  1. The theoretical results do not cover the deployed operator, nor do they explain the one-step refinement used in practice. Theorem 2.1 and Proposition 2.2 analyze the standard, unnormalized Hopfield update, whereas HoReN deploys a normalized, damped variant with intermediate re-normalization and a subsequent normalization step, typically executed only once. At M=1M=1, Proposition 2.2 provides only a fixed-point residual bound of 2\sqrt{2}, rather than a performance guarantee for the deployed update. Since both the initial query and the updated query have norm at most one, their difference is in any case bounded by two; the proposition therefore offers limited insight into why the particular one-step rule improves retrieval. The theoretical explanation that multi-step iteration also attracts unrelated queries toward codebook-induced fixed points can motivate the design, but it does not theoretically validate the effectiveness of the deployed one-step algorithm.

  2. Simpler query-refinement baselines are missing, so the necessity of the full Hopfield update remains unproven. With β=20\beta=20, the softmax distribution will often be sharply concentrated on a small number of nearest keys. Consequently, the current one-step Hopfield refinement may effectively amount to interpolating the query toward the top-1 key, or to increasing the confidence of an existing nearest-neighbor match. The paper should compare against the following simpler mechanisms: directly interpolating the normalized query toward the top-1 key; applying temperature calibration or a monotonic transformation to the top-1 cosine score; using a weighted centroid of the top-kk keys instead of a softmax over the entire codebook; and performing nearest-neighbor matching alone with the same threshold calibration.

  3. The claim that “adding entries does not move existing decision boundaries” is not valid, and scale stability remains primarily an empirical observation. Section 3.2.2 partially attributes scale stability to the claim that adding entries does not move existing decision boundaries. However, both p=softmax⁡(βqK⊤)p=\operatorname{softmax}(\beta q K^\top) in the Hopfield update and the final arg⁡max⁡iqki⊤\arg\max_i qk_i^\top depend on the entire codebook. Adding a new key can change the softmax weights, the refined query, and the nearest-neighbor competition, thereby changing the effective decision regions of existing keys. Fixed stored keys therefore do not imply fixed routing boundaries.

  4. Baseline calibration is not fully symmetric, and some of the extreme baseline failures require stronger fairness checks. Appendix E.4.1 explains that the baselines largely reuse hyperparameters from their original papers, with only module-name and model-family-level adaptations when transferred to models not studied in those papers. In contrast, HoReN selects its matching threshold and token pooling ratio for each model. Although using the original recipes is a reasonable starting point, GRACE’s Generalization remains only 0.01–0.06 on almost all models under this setup, while WISE also approaches near-total collapse on some models. The authors should rule out systematic disadvantages caused by thresholds, distance scales, layer selection, or implementation adaptation before attributing these failures entirely to the baseline methods.

If the authors can provide systematic comparisons against simpler alternatives such as top-1 key interpolation, a top-kk weighted centroid, and score calibration, I would consider raising my score.

Questions for authors
  1. How does the simple baseline of “normalized query + interpolation toward the top-1 key with the same step size” compare with full Hopfield refinement? Can any remaining difference demonstrate the necessity of the global softmax attractor field?
  2. After recalibrating key parameters of GRACE and WISE—such as thresholds and distance scales—under the same validation protocol, do the paper’s main performance gaps remain unchanged?
  3. At 10K and 50K edits, what are the retrieval success rate on paraphrased edit queries, the false-positive rate on unrelated queries, and the similarity margin between the target key and its nearest competing key?
  4. At 50K edits, what are the actual codebook storage cost, per-query latency, and total editing time? Do Reliability, Generalization, and Locality remain stable when an approximate nearest-neighbor index is used?
  5. When successive conflicting updates arrive for the same subject and relation, how does HoReN ensure that it returns the most recent fact rather than allowing similarity-based competition between the old and new codebook entries to determine the answer?

Comments on the paper

1 comment

3. Why the refinement rule can also affect memory construction.

Refinement is applied before allocation/reuse decisions during editing, not only at inference. A query pushed above the acceptance threshold can reuse a same-label entry, while a different refinement direction can keep it below the threshold and cause a new key to be stored. The resulting coverage can therefore differ even though both methods store original, unrefined keys.

We examined this mechanism on a selected 100-example sequence with fixed edit order. Across five additional initializations, the mean results were:

MethodRel (%)Gen (%)Loc (%)OP (%)Learned keys, excluding sentinel
HoReN99.9595.2610098.3899
Top-1 interpolation98.9589.2610095.9588

Each run had seven HoReN wins, one Top-1 win, and 92 ties. On the seven HoReN-winning paraphrases, swapping the memories gave:

Memory usedHoReN router: correct cases out of sevenTop-1 router: correct cases out of seven
HoReN77
Top-100

For example, editing “What is the sex of Soraia André?” with target “female” found an existing same-label entry for Phyllis Chan. HoReN scored 0.847656, below 0.85, and added a key; Top-1 scored 0.867188 and reused the entry. On the later paraphrase, HoReN retrieved its new key at 0.945312 and answered “female,” whereas Top-1 accepted an unrelated entry at 0.851562 and answered “French.”

The swap localizes the seven-case difference to the constructed memory, rather than to an inference-router advantage on the same memory. HoReN's better coverage here also costs more entries. The repeated initializations concern this selected sequence, not independent streams. The case study uses a float32 threshold comparison, while the default-stream controls use the legacy BF16 gate; one of the seven allocation differences is precision-sensitive, whereas the Soraia split persists in higher-precision geometry replay.

4. Interpretation and manuscript revision.

The shared-threshold comparison characterizes the Gen–Loc trade-off against a competitive Top-1 alternative. The memory-swap study shows a concrete effect on key allocation and paraphrase coverage beyond an inference-time confidence change. The direct Top-5 control additionally characterizes centroid replacement, with its distinct update strength. Together, these results support a role for refinement in the tested editor, while the necessity of full-codebook rather than matched-damping Top-k refinement remains unresolved. We will revise Section 3.4 and Figure 4 to separate complete-editor gains from the independent contribution of refinement, and present the allocation study as a mechanism case study with its explicit memory cost.

Point 3 (Theory) Decision boundaries under codebook growth

**Weakness W3 **

The claim that “adding entries does not move existing decision boundaries” is not valid, and scale stability remains primarily an empirical observation. Section 3.2.2 partially attributes scale stability to the claim that adding entries does not move existing decision boundaries. However, both p=softmax⁡(βqK⊤)p=\operatorname{softmax}(\beta qK^{\top }) in the Hopfield update and the final arg⁡max⁡iqki⊤\arg \max _{i}qk_{i}^{\top } depend on the entire codebook. Adding a new key can change the softmax weights, the refined query, and the nearest-neighbor competition, thereby changing the effective decision regions of existing keys. Fixed stored keys therefore do not imply fixed routing boundaries.

Response

Thank you for identifying this issue. We agree that fixed stored keys do not imply fixed effective routing boundaries. We will correct the sentence in Section 3.2.2 and distinguish the following two concepts.

1. What remains fixed: the score equations of existing keys.

Let zz denote a unit-normalized refined query. For fixed existing key vectors and threshold cc, the equations

\qquad z^\top(k_i-k_j)=0$$ describe an existing key's threshold surface and an existing pair's equal-score surface. Their equations remain fixed, while their role in the routing partition depends on the other keys. **2. What can change: both the query refinement and the final competition.** Write $z_K(x)$ for the refined query produced from input $x$ using codebook $K$. The input region routed to key $i$ is determined by - **Acceptance:** $z_{K}(x)^{\top}k_{i}>c$. - **Highest score:** $z_{K}(x)^{\top}k_{i}\geq z_{K}(x)^{\top}k_{j}$ for every key $j$. Equal scores are resolved using the implementation's tie-breaking rule. Adding a new key can change this region through two mechanisms: - **Refinement:** the new key participates in the softmax-weighted memory read, so the same input can produce a different $z_K(x)$. - **Competition:** new key introduces an additional candidate in the final argmax and can replace the previous winner. This mechanism exists even when refinement is disabled.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.