acceptodds
Under review as a conference paper at ICLR 2027

KGHEdit: Knowledge-Graph-Guided Continual Model Editing via Heat Diffusion

Abstract

Large language models are increasingly deployed in domains with continuously changing knowledge, making post-pretraining updates necessary without repeated full-model retraining or loss of broad model capabilities. Model editing offers a practical path for such deployment-time updates because it can modify parametric knowledge without retraining the full model. However, existing editors remain poorly matched to continual domain updating: they typically treat each update as an independent fact rewrite, even though knowledge is structured and interconnected, and correlated edits can interact through shared entities and relations. We propose **KGHEdit**, a knowledge-graph-guided framework for continual knowledge editing that models incoming edits as local heat sources and uses heat diffusion to identify and weight structurally related support for a base editor. Together, matched controls, interaction-focused streams, and a 10K-edit audit provide convergent evidence that graph-guided support selection and conditional activation offer a practical path to more reliable continual knowledge updating.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
knowledge graph embedding, 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 4 reviews, above 27% of the others.
Reviewer_uGjGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

KGHEdit augments a continual parameter-editing method with support examples selected from a local knowledge graph. It treats the current edit and prior edits as heat sources, scores nearby entities by damped graph diffusion, and filters candidate examples using compatibility and conflict-risk rules before passing a weighted batch to alphaedit or NAS. The paper reports matched editing results on wikibigedit, constructed medical streams, and a 10,000-edit counterfact evaluation. It also compares heat diffusion with alternative support selectors and presents ablations of its balance component. The central claim is that graph-guided scope selection and historical balancing improve continual editing while preserving unrelated knowledge.

Critical strengths
  1. The method improves several matched editing outcomes. Table 1 compares KGHEdit with the same base editor under matched backbone, request order, and stream length. The NAS gains on Llama, Qwen, and GPT-J make the empirical result more consequential than a single-editor demonstration.
  2. Support selection is tested against meaningful alternatives. Table 5 holds the editor and support budget fixed while comparing diffusion with k-hop, entity retrieval, PPR, and a no-heat variant. This addresses the immediate concern that any larger support batch would yield the same result.
Critical weaknesses
  1. The method improves several matched editing outcomes. Table 1 compares KGHEdit with the same base editor under matched backbone, request order, and stream length. The NAS gains on Llama, Qwen, and GPT-J make the empirical result more consequential than a single-editor demonstration.
  2. Support selection is tested against meaningful alternatives. Table 5 holds the editor and support budget fixed while comparing diffusion with k-hop, entity retrieval, PPR, and a no-heat variant. This addresses the immediate concern that any larger support batch would yield the same result.
Questions for authors
  1. What is the substantive continual-editing advance over GLAME beyond using a different graph-based support score? Which part of the reported gain do the authors attribute to historical coupling?
  2. How do the authors reconcile the no-balance result in table 5 with the identical alpha=0alpha=0 and alpha=0.2alpha=0.2 scores in appendix table 12? In the main streams, when does memory change the selected support or the edited model’s answers?
  3. What graph source and snapshot were available at edit time for each evaluation? Could graph-derived support contain facts from later requests or answers to propagation and locality probes?
3 comments
AuthorsOpenReview ↗

Response to Reviewer uGjG

Thank you for recognizing the matched improvements across base editors and the controlled support-selection comparisons. We appreciate your questions about the continual-editing advance, the role of historical information, and the information available during editing.

Q1. What does KGHEdit add beyond GLAME, and what is the role of history? (Q1)

KGHEdit maintains a decayed graph footprint of previous edits and combines this footprint with the current edit's local heat source (Section 3.3). Before each parameter update, it uses compatibility and collision-risk checks to construct a bounded support batch (Section 3.4). This stateful support-planning process allows earlier edits to influence later support selection when their graph neighborhoods interact.

We added a three-seed, same-stream Fixed-4 comparison with GLAME. The methods use matched request streams and the same underlying update interface; the comparison tracks behavior as edits accumulate:

EditsGLAMEKGHEdit
200.76530.6893
400.77270.7125
600.39050.7151
1000.25700.7190

GLAME performs better at the early checkpoints, while KGHEdit retains higher performance at later checkpoints; the GLAME runs also show substantial variation across seeds at longer horizons. This comparison illustrates the long-stream behavior of the two support-planning approaches.

To examine when historical interactions matter, we separately evaluated the pre-edit gate on three-seed natural temporal-repeat streams. Enabling the gate improves later repeated-fact performance by +0.0166 (95% CI [+0.0039, +0.0283]). The GLAME comparison and the gate experiment thus provide complementary evidence about long-horizon behavior and conditional support activation, respectively.

Q2. Why do No-Balance and the α sweep differ, and when does history change the outcome? (Q2)

The two experiments intervene on different operations. In Section 3.3, α controls the contribution of previous-edit memory to the diffusion source. In Section 3.4, Balance applies compatibility and collision-risk checks when the heat-ranked candidates are compiled into the edit batch.

Table 5 shows that disabling Balance reduces Propagation from 0.2792 to 0.2375 on its matched proxy stream. Appendix Table 12 varies α on a separate 96-edit stress stream, where α=0 and α=0.2 both yield 0.3356 Propagation. The former measures the support compiler; the latter measures source mixing under that stress-flow configuration. We will make this distinction explicit beside both results.

We also measured the interaction-dependent effect of activating auxiliary support in natural temporal-repeat streams:

Later-repeat measureGainPaired-case bootstrap 95% CI
Gate+0.0166[+0.0039, +0.0283]
Support contribution+0.0183[+0.0021, +0.0542]

Both gains are positive on the later-repeat subset. In the separate frozen single-seed Heat × History-source intervention, the source toggle does not change the final score and changes selected support in one Heat round. These observations identify the repeated-slot activation regime where support affects measured outcomes, while keeping the effects of the α source and the Balance compiler distinct. We will add this distinction to the mechanism discussion and align the terminology across the main text and appendix.

Q3. Could KG support contain future requests or evaluation answers? (Q3)

KGHEdit uses two information sources at edit time: the external KG neighborhood linked to the current request, and planner memory containing previously submitted edits. Algorithm 1 updates planner memory only after the current parameter write. Appendix Table 8 separately reserves Rephrase, Persona, Locality, and Propagation prompts and answers for evaluation, so those evaluator fields are not supplied directly to the support selector.

The external KG is a distinct source of factual information. To establish its temporal and factual boundary, the relevant checks are the KG source and snapshot for each stream, together with fact-level comparisons between selected support, later edit targets, and evaluation answers. We will specify the KG provenance and the available overlap checks explicitly in the experimental details. The current protocol documents the field-level separation; a completed triple-level overlap audit is not included in the reported evidence.

We appreciate your recognition of the matched editing results and your focus on the mechanism behind continual gains. We hope the new GLAME comparison, repeated-fact analysis, and clearer information-boundary description help resolve your questions, and we would be grateful if you would reconsider your rating in light of the additional evidence.

AuthorsOpenReview ↗

Follow-up on Continual Editing and Historical Support

Dear Reviewer uGjG,

Thank you again for your constructive feedback and recognition of our experimental results. In response to your questions, we added a matched comparison with GLAME, examined when historical support improves editing through repeated-fact experiments, and clarified the distinction between memory-based diffusion and Balance. We also explained the information available at edit time and its separation from evaluation inputs.

As the ICLR discussion period is approaching its end, we would sincerely appreciate your thoughts on these responses and welcome any further discussion of your remaining questions.

Thank you very much for your time and consideration.

AuthorsOpenReview ↗

KGHEdit (#17368): Request for Response from Reviewer uGjG

Dear Reviewer uGjG,

As the discussion period is nearing its end, we sincerely hope to hear your thoughts on our rebuttal.

We have added a direct comparison with GLAME, examined the contribution of historical support, and clarified the information available during editing.

We would greatly appreciate it if you could review these results and let us know whether your main concerns have been addressed.

Thank you very much for your time and consideration.

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

This paper proposes KGHEdit, a knowledge-graph-guided framework for continual model editing. It uses heat diffusion over local KG neighborhoods, together with historical edit memory and compatibility filtering, to select auxiliary examples for a base editor. Experiments across multiple models and benchmarks, examine its effectiveness in improving editing performance and knowledge preservation.

Critical strengths
  1. The paper is well organized and easy to follow, with illustrative examples and a clear presentation of the framework.
  2. The motivation is meaningful: continual updates can interact through related facts, making edit scope and consistency important beyond individual rewrite success.
  3. The proposed approach is conceptually reasonable and modular. Combining graph-based relevance, historical support, and conflict filtering offers a plausible way to improve existing editors without redesigning their parameter-update mechanisms.
Critical weaknesses
  1. The ablations do not sufficiently isolate the core mechanisms. In Table 5, removing balance reduces Propagation by 0.0417, compared with 0.0148 when replacing heat diffusion with k-hop retrieval. The four non-heat alternatives cluster within 0.2642–0.2664, showing no clear benefit from graph structure in these controls. Moreover, PPR also performs diffusion-based smoothing, so “smooth ranking versus flat retrieval” does not explain its gap from the proposed operator. Please define the balance switch precisely, complete the factorial ablation, and validate the findings under the main experimental protocol, controlling for normalization and support-selection budgets.
  2. The paper identifies local KG triples and historical edits as candidate sources, but does not adequately explain how they become prompt–target pairs or how outdated targets are handled. A complete example and an overlap analysis between auxiliary training facts and Propagation evaluation facts would help distinguish direct incorporation of supplied facts from generalization to facts not explicitly provided.
  3. The method depends on entity linking, graph coverage, and factual freshness, yet these conditions are not systematically varied. Entity-linking perturbations, edge deletion, and a stale-KG setting would clarify when the method remains beneficial and whether incorrect or outdated support causes negative transfer relative to the base editor.
Questions for authors
  1. Diffusion converts directed triples into undirected unit-weight edges, retaining relation labels only for downstream filtering. Please justify this simplification and assess whether preserving direction or relation type changes support selection and propagation performance.
  2. Under §3.4, support is activated only when the current request overlaps a previously observed subject–relation slot. Structurally related edits, such as two different companies sharing a parent organization, need not satisfy this condition. Please clarify how such cases are handled, report activation rates across interaction types, and confirm that the same gate is used in the main experiments, ablations, and case studies.
4 comments
AuthorsOpenReview ↗

Response to Reviewer zcij

Thank you for your careful analysis of the component ablations, support construction, KG quality, and activation rule. Your questions help us clarify how graph scoring, historical sources, and support compilation contribute to continual editing.

Q1. How are the effects of Heat and Balance separated? (W1)

KGHEdit uses Heat to rank structurally related candidates and Balance to compile support using compatibility and conflict-risk checks (Sections 3.2–3.4). Table 5 compares these mechanisms under the same editor, request stream, evaluator, and support budget. Replacing Heat while retaining Balance tests the ranking rule; disabling Balance while retaining Heat tests the support compiler.

ConfigurationUpdateLocalityPropagation
Heat + Balance0.78360.43910.2792
k-hop + Balance0.76620.41660.2644
PPR + Balance0.75870.42160.2662
Heat + No-Balance0.77400.41970.2375

These matched contrasts show complementary effects: Heat improves the selection of relevant support within a local neighborhood, while Balance contributes more strongly to propagation by screening competing support. The comparison with PPR is especially informative because both methods smooth information over a graph, yet use different scoring operators.

We also evaluated a frozen 320-request Heat × History-source intervention with fixed support conditions. In this single-seed test, Heat improves Propagation over k-hop (0.38902 versus 0.36416), while changing the historical source leaves the final metric unchanged. The historical mixing coefficient α in Section 3.3 controls the source supplied to diffusion; the Balance switch controls subsequent support compilation. We will distinguish these interventions explicitly and identify the full Heat × Balance factorial as a separate evaluation.

Q2. How are KG triples converted into support, and how is evaluation information separated? (W2)

For each incoming edit, KGHEdit first links the request entities to a local KG neighborhood (Section 3.1). It then collects candidate triples and eligible previous-edit records, computes their heat, compatibility, and risk scores using Eq. (3), and selects a bounded support set. Each selected fact is converted into a prompt–target pair with a normalized weight and added to the target edit batch before the base editor writes parameters (Section 3.4; Appendix B). This sequence makes the source and role of every auxiliary example explicit.

Appendix Table 8 specifies the information available at edit time: target request fields, local KG triples, and prior submitted edit records. Rephrase, Persona, Locality, and Propagation prompts and answers are reserved for evaluation. This establishes the field-level separation of editing and evaluation inputs. We will additionally provide a concrete triple-to-prompt example and document the KG provenance and snapshot used in each protocol. Fact-level comparison of committed support with future targets and evaluation facts is a distinct audit needed to characterize possible semantic overlap; we will state the scope of the available checks precisely.

Q3. What happens under graph corruption? (W3)

We added a 39-condition audit covering edge deletion, missing nodes, and relation mislinking. As graph coverage decreases, the support selector reduces the number of eligible auxiliary facts; when no eligible neighbors remain, the edit batch falls back to the target request. This provides a controlled response to incomplete graph neighborhoods.

A matched single-seed comparison on the same 320-request stream illustrates the effect on the edited model:

ConditionOfficialPropagation
Target-only0.622660.34321
Clean KGHEdit0.631560.38902
30% edge deletion0.625720.37214

Propagation remains above the target-only result after 30% edge deletion in this setting. The broader perturbation matrix also measures how incorrect links alter the selected support. Together, these analyses characterize the behavior of KGHEdit as graph coverage and correctness vary; we will present the corruption protocol and fallback rule alongside the results.

Q4. Why use an undirected unit-weight graph? (Q1)

We symmetrize each directed KG triple for the diffusion operator, while retaining relation multiplicity in the adjacency matrix (Section 3.1). This produces a nonnegative symmetric Laplacian; damping makes the linear system strictly positive definite, yielding a unique and stable heat field (Proposition 1). The heat field estimates structural reach, and the subsequent Compat and Risk terms preserve relation identifiers, target types, and subject–relation conflicts during support selection (Eq. (3); Appendix Table 7).

The design therefore combines a stable graph influence score with relation-aware support compilation. We will clarify this separation in the method description. Direction-preserving and relation-weighted diffusion are useful extensions, and the present experiments do not include a direct comparison of those operators.

Q5. When does the pre-edit gate activate? (Q2)

The pre-edit gate in Section 3.4 activates auxiliary support when the current request revisits a previously observed subject–relation slot. This criterion identifies interactions where an earlier update can directly affect a later request. When the gate is inactive, the base editor receives the target request alone.

We tested this rule on frozen natural temporal-repeat streams with 320 requests per seed and three seeds:

SubsetGate gainPaired-case bootstrap 95% CI
Later repeated facts+0.0166[+0.0039, +0.0283]
Singleton facts+0.0013[-0.0087, +0.0166]
All requests+0.0023[-0.0066, +0.0162]

The later-repeat gain is positive in all three seeds, providing direct evidence for conditional support activation in the repeated-slot regime. Two requests that share a graph neighbor but have different subject–relation slots do not activate this particular gate. Their structural relationship can still be represented in the local KG, and the current gate reserves auxiliary editing for the more directly observable interaction condition. We will state this activation rule consistently in the method, experimental protocols, and case descriptions.

We appreciate the precision of your questions, particularly the distinction between graph scoring, support compilation, and evaluation information boundaries. We hope the matched ablations, new corruption tests, and explicit activation analysis clarify the method's contributions, and we would welcome your reconsideration of the score based on these clarifications.

AuthorsOpenReview ↗

Follow-up on Mechanism Analysis and Additional Experiments

Dear Reviewer zcij,

Thank you again for your careful review and detailed suggestions. Following your feedback, we clarified the separate roles of Heat and Balance, added experiments on graph corruption and conditional support activation, and explained how KG facts are converted into editing support. We also clarified the information available during editing and the scope of our evaluation.

As the ICLR discussion period is approaching its end, we would greatly appreciate your thoughts on these additional results and clarifications. We would welcome the opportunity to discuss any remaining concerns.

Thank you very much for your time and consideration.

AuthorsOpenReview ↗

KGHEdit (#17368): Request for Response from Reviewer zcij

Dear Reviewer zcij,

As the discussion period is approaching its end, we sincerely hope to have the opportunity to discuss our rebuttal with you.

Following your suggestions, we have provided further analysis of Heat and Balance, graph corruption, support construction, and conditional activation.

We would be very grateful if you could take a moment to review these responses and share any remaining concerns or feedback.

Thank you very much for your time and consideration.

Reviewer_xoK5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work addresses the topic of knowledge edit by proposing a structure-aware framework that models the batch of knowledge update as heat diffusion. The experiments are mainly based on the WikiBigEdit benchmark and the results show the improvements made by the proposed method on Llama 3.1 8B, Qwen 2.5 7B, and GPT-J 6B, in several cases. Efficiency of the proposed method was also measured and compared.

Critical strengths
  • This paper is clearly written and easy to follow.
  • Experiments show the several improvements made by the proposed method.
  • The proposed method can be integrated with other popular approaches such as AlphaEdit.
Critical weaknesses
  • The experiments did not include more recent LLMs. In particular, none of the reasoning LLMs were evaluated in this work. As reasoning LLMs have been the standard since 2025, the effectiveness of the proposed method for current models is still unclear.
  • As reported by related work, knowledge edit is risky to degrade the LLM’s general, broad ability, in particular after a sequence of edits. This work analyzes the locality trade-off only, but not include an analysis on the trade-off in general abilities. Without this kind of evaluation, the practical usability of the proposed method is still unclear. The general benchmarks like BBH and MMLU-Pro could be adopted to measure the general abilities of the target LLM along the sequence of knowledge edits.
Questions for authors
  • Can the proposed method be applied with a reasoning LLM?
3 comments
AuthorsOpenReview ↗

Response to Reviewer xoK5

Thank you for your constructive suggestions concerning reasoning-tuned models and the preservation of general capabilities. We appreciate the opportunity to clarify KGHEdit's editing interface and provide additional capability measurements.

Q1. Can KGHEdit be applied to reasoning LLMs? (W1 & Q1)

KGHEdit constructs a weighted prompt–target batch before the base editor updates model parameters (Section 3). Its local KG construction, heat scoring, and support selection operate on factual requests and their relations, independently of the model's internal reasoning procedure. Consequently, the framework can be integrated with a reasoning-tuned backbone whenever its parameter-writing editor accepts the same batch interface.

Our current evaluations establish compatibility across three backbone families and show that this graph-side interface can improve editing beyond one model architecture. For example, with Qwen2.5-7B and NAS:

MetricNASNAS + KGHEditGain
Personas0.51050.6548+0.1443
Propagation0.25730.3198+0.0625

These results support the portability of the editing interface. Evaluation on a reasoning-tuned checkpoint, including reasoning retention after sequential edits, remains an important next validation step. We will make this distinction and the compatibility requirements explicit in the paper.

Q2. What happens to broad capabilities after continual editing? (W2)

We agree that preservation should be evaluated beyond factual Locality. We therefore added three-seed MMLU and CommonsenseQA evaluations under matched editing conditions, comparing KGHEdit with the target-only editor after continual editing:

BenchmarkKGHEdit − Target-only
MMLU-0.0072
CommonsenseQA-0.0022
Combined-0.0037

The measured differences are small in absolute terms. CommonsenseQA and the combined evaluation satisfy the prespecified non-inferiority criteria; the MMLU-specific criterion is not met. These results provide an independent capability check alongside the editing improvements in Table 1 and make the observed trade-offs visible at the benchmark level. We will report the benchmark-specific outcomes and evaluation protocol clearly, and further evaluate reasoning-oriented capabilities on suitable checkpoints.

We appreciate your emphasis on practical model compatibility and general capabilities. We hope these additional measurements and the clarified scope of the editing interface address your main concerns, and we would be grateful if you would reconsider your assessment in light of this evidence.

AuthorsOpenReview ↗

Follow-up on Model Compatibility and Capability Preservation

Dear Reviewer xoK5,

Thank you again for your thoughtful review and helpful suggestions. In our rebuttal, we clarified how KGHEdit can be integrated with different model backbones through a compatible editing interface, and added MMLU and CommonsenseQA evaluations to examine general capability preservation after continual editing. We also clarified the current scope and limitations regarding reasoning-tuned models.

As the ICLR discussion period is approaching its end, we would sincerely appreciate your thoughts on whether these additions address your concerns. We would be happy to discuss any remaining questions.

Thank you very much for your time and consideration.

AuthorsOpenReview ↗

KGHEdit (#17368): Request for Response from Reviewer xoK5

Dear Reviewer xoK5,

As the discussion period is nearing its end, we sincerely hope to hear your thoughts on our rebuttal.

We have clarified the applicability of KGHEdit to different model backbones and provided additional MMLU and CommonsenseQA experiments to examine general capability preservation.

We would greatly appreciate it if you could review our responses and let us know whether they address your concerns.

Thank you very much for your time and consideration.

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

This paper proposes KGHEdit, a knowledge-graph-guided framework for continual knowledge editing that aims to improve consistency across related factual updates while preserving unrelated knowledge. The method constructs a local knowledge graph around each edit, applies damped heat diffusion to estimate the relevance of neighboring entities, and incorporates historical edits through a decayed entity-memory mechanism. The resulting heat scores, together with compatibility and conflict-risk heuristics, determine a weighted set of supporting examples for an existing base editor, such as AlphaEdit or NAS. Experiments across WikiBigEdit, a constructed medical editing benchmark, and CounterFact-10K demonstrate improvements in several editing and preservation metrics. The paper further investigates the framework through component ablations, trajectory analyses, mechanism audits, and efficiency measurements.

Critical strengths

Strengths

  1. The paper addresses an important problem in continual knowledge editing. It highlights the limitations of treating related factual updates independently and proposes a structure-aware approach to improve editing consistency and reduce interference.
  2. KGHEdit provides a flexible framework for existing editing methods. It combines knowledge-graph diffusion, historical edit memory, and conflict-aware support selection, allowing integration with different base editors without redesigning parameter-update mechanisms.
  3. The paper provides extensive experiments and useful analyses. Evaluations cover multiple models, editors, domains, and up to 10,000 edits, showing encouraging improvements. Ablations and trajectory analyses further investigate the contributions of different components.
Critical weaknesses
  1. The improvements are mixed across editors and metrics. In Table 1, KGHEdit reduces both Update and Locality when applied to AlphaEdit on Llama-3.1-8B, UltraEdit outperforms all KGHEdit variants on Locality and Propagation at N2000. The paper lacks sufficient analysis of these trade-offs and when graph-guided support may negatively affect editing.
  2. The experiments do not fully isolate the contribution of heat diffusion. Although Table 5 compares alternative support-selection strategies under matched budgets, these controls are limited to a proxy stream rather than the main long-horizon experiments. The experiments also did not compare against cited closely related KG-aware method GLAME.
  3. The method relies on knowledge-graph quality without sufficient robustness evaluation. KGHEdit depends on accurate entity linking and graph structure, but the experiments do not systematically examine incomplete, noisy, or outdated graphs.
  4. The theoretical analysis is not sufficiently connected to the practical editing improvements. Propositions 1 and 2 restate standard properties of a positive-definite linear system. Proposition 3 analyzes a coupled objective that KGHEdit never solves, since the implementation uses a heuristic risk score.
Questions for authors
  1. How is the knowledge graph constructed and updated throughout the continual editing? Especially if there are outdated triples or contradictions.
  2. How much of KGHEdit's improvement comes specifically from heat diffusion rather than other components like graph based retrieval?
4 comments
AuthorsOpenReview ↗

Response to Reviewer eHEv

Thank you for recognizing the importance of structured continual editing and the breadth of our evaluation. We address your concerns about editor-dependent effects, component attribution, KG quality, and the theoretical connection.


Q1. When does KGHEdit help, and how does its effect depend on the base editor? (W1)

You are right to highlight the variation across editors. KGHEdit is particularly effective with NAS over long streams, where repeated parameter updates accumulate. The matched results in Table 1 show sizeable gains in rewriting and generalization, including on a second backbone.

Backbone / editorEditsUpdateRephrasePersonasPropagation
Llama / NAS2,000+0.0486+0.0366+0.0369+0.0009
Llama / NAS3,000+0.0418+0.0320+0.0271+0.0319
Qwen / NAS1,000+0.0409+0.0513+0.1443+0.0625

All entries are KGHEdit minus the matched base editor. With AlphaEdit at 2,000 edits, Propagation improves by 0.0222 and Rephrase by 0.0039, while Update and Locality decrease by 0.0172 and 0.0147. The results identify a meaningful editor-dependent trade-off: the strongest multi-axis benefits occur with NAS, while AlphaEdit favors propagation at some cost to rewriting and preservation. This is why our framework retains the underlying editor and makes support selection an explicit, controllable stage.


Q2. What is the contribution of heat diffusion beyond graph retrieval? (W2 & Q2)

We agree that the selection rule should be tested under a common support budget. Table 5 holds the editor, stream, evaluator, and support budget fixed. Heat gives the strongest Update, Rephrase, and Propagation among the tested scope rules, including PPR, which also performs diffusion-style smoothing.

Scope ruleUpdateRephrasePropagation
Heat0.78360.68920.2792
k-hop0.76620.66990.2644
PPR0.75870.65840.2662
Entity retrieval0.75190.65430.2642

The damped Laplacian scores candidate entities according to their paths and local graph connectivity, and the compatibility/risk compiler determines which scored facts enter the write batch. This separation makes the contribution of heat-based ranking directly testable. These Table 5 comparisons use the strict proxy stream; our longer WikiBigEdit results establish the performance of the full framework under the main editing protocol.


Q3. What is the advance over GLAME? (W2)

We appreciate your request for a direct comparison. GLAME uses KG-associated facts to strengthen edits. KGHEdit additionally maintains decayed footprints of earlier edits, combines them with the current graph source, and screens compatibility and collisions before writing parameters. In a same-stream, three-seed Fixed-4 comparison, KGHEdit maintains its later-checkpoint score as the edit sequence grows.

EditsGLAMEKGHEdit
200.76530.6893
400.77270.7125
600.39050.7151
800.38940.7180
1000.25700.7190

The values are three-seed means under the reported Fixed-4 setting. GLAME has substantial variation across later seeds, so we interpret this comparison within that protocol. A separate fixed-update plugin check holds the GLAME update rule and support count constant while changing selected support identities, establishing a compatible intervention interface.


Q4. How does KG quality affect the method? (W3)

We evaluated edge deletion, missing nodes, and relation mislinking in 39 audited perturbation cells. As edges or nodes disappear, the selector reduces its support count and falls back to the target-only edit when no support remains. In a matched single-seed direct test, Propagation changes as follows:

KG conditionPropagation
Target-only0.34321
Clean KG0.38902
30% edge deletion0.37214

This illustrates the controlled reduction of graph-derived support under incomplete coverage. The broader perturbation matrix also measures incorrect relation links; its estimates vary around the clean-graph result, so the operational conclusion is the explicit support-selection and fallback behavior rather than a uniform improvement under all corruptions.

Q5. How does the theory inform the implemented editor? (W4)

Proposition 1 establishes a unique, stable damped diffusion field; Proposition 2 gives its exact linear decomposition into current and historical sources. Proposition 3 shows that a coupled conflict penalty attenuates incompatible overlap locally. Our implementation uses the corresponding Heat, Compat, and Risk terms in Eq. (3), avoiding pairwise coupled solves. Theorem 1 then bounds the realized one-step response under the stated local first-order editor assumption. These results motivate the planner's structural scoring and conflict screening, while the matched editing experiments measure their downstream effect.


Q6. How are outdated or conflicting KG facts handled? (Q1)

For each request, we form a local neighborhood from the available external KG and maintain a separate decayed record of previously submitted edits (Algorithm 1). Candidates are checked for relation/target compatibility and same-slot conflicts before they enter the weighted batch (Eq. (3), Appendix Table 7). The current target remains in the batch when auxiliary support is filtered. The external KG source and its temporal snapshot determine which triples are available; the submitted experiments do not establish a general stale-triple replacement guarantee. We will identify that provenance explicitly in the experimental description.

AuthorsOpenReview ↗

Follow-up on Rebuttal and Further Discussion

Dear Reviewer eHEv,

Thank you again for your detailed and constructive feedback. Following your suggestions, we clarified the editor-dependent trade-offs, added a direct comparison with GLAME, and examined the effects of graph corruption. We also provided matched comparisons to better isolate the contribution of heat diffusion and clarified its connection to the theoretical analysis.

As the ICLR discussion period is approaching its end, we would be grateful to hear your thoughts on these additional results and whether they help address your main concerns. We would welcome any further questions or discussion.

Thank you very much for your time and consideration.

AuthorsOpenReview ↗

KGHEdit (#17368): Request for Response from Reviewer eHEv

Dear Reviewer eHEv,

As the discussion period is approaching its end, we sincerely hope to have the opportunity to discuss our rebuttal with you.

We have provided additional experiments addressing your concerns about heat diffusion, comparisons with GLAME, and robustness to imperfect knowledge graphs.

We would be very grateful if you could review our responses and share your thoughts on whether they address your main concerns.

Thank you very much for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary: Key Revisions and Additional Evidence

We sincerely thank all reviewers for their thoughtful and constructive feedback. In response, we conducted additional experiments and clarified the main mechanisms, evaluation protocols, and limitations of KGHEdit. The key updates are summarized below.

  1. Clarifying when KGHEdit improves continual editing. We analyzed the differences across base editors and model backbones. KGHEdit shows consistent improvements with NAS, including gains in Update, Rephrase, and Personas on Llama and Qwen. We also clarified the trade-offs observed with AlphaEdit, where improvements in Propagation may come with reduced Update or Locality performance.

  2. Isolating the effects of Heat and Balance. Under matched editing and support budgets, Heat achieves better Update (0.7836) and Propagation (0.2792) than k-hop (0.7662/0.2644), PPR (0.7587/0.2662), and entity retrieval (0.7519/0.2642). Disabling Balance reduces Propagation from 0.2792 to 0.2375. These results help distinguish heat-based ranking from compatibility and conflict-aware support selection.

  3. Comparing with GLAME. We added a matched three-seed comparison using the same Fixed-4 protocol. While GLAME performs better at early checkpoints, KGHEdit maintains higher scores later in the editing stream. At 100 edits, KGHEdit achieves 0.7190 compared with 0.2570 for GLAME. This provides additional evidence of KGHEdit's behavior under accumulating edits.

  4. Examining historical support and conditional activation. We clarified that historical memory and Balance perform different operations. On natural temporal-repeat streams, conditional support activation improves later repeated-fact performance by 0.0166 (95% CI [0.0039, 0.0283]), while the improvement on singleton facts is much smaller. These findings help identify when historical interactions are most useful.

  5. Evaluating robustness to imperfect knowledge graphs. We added 39 graph-perturbation conditions covering edge deletion, missing nodes, and relation mislinking. In a matched single-seed test, Propagation reaches 0.3890 with a clean KG and 0.3721 after 30% edge deletion, compared with 0.3432 for target-only editing. We also clarified the fallback behavior when graph support becomes unavailable.

  6. Clarifying information boundaries and graph construction. We explained how KG triples and historical edits are converted into weighted prompt–target support, and how the pre-edit gate determines when support is activated. We also clarified that evaluation prompts and answers are not directly supplied to the support selector. The external KG source and snapshot remain important for interpreting possible factual overlap, and a complete triple-level overlap audit is not included in the current evidence.

  7. Model compatibility and general capability preservation. We clarified that KGHEdit operates through a model-independent support-selection interface and can be integrated with different compatible base editors. Experiments across Llama, Qwen, and GPT-J demonstrate its applicability to multiple backbone families. We further added three-seed MMLU and CommonsenseQA evaluations to assess general capability preservation after continual editing. The results show small average changes relative to target-only editing (-0.0072 on MMLU and -0.0022 on CommonsenseQA), providing additional evidence of controlled effects on broader model capabilities.

  8. Strengthening the theoretical interpretation. We clarified how the diffusion analysis motivates structural scoring, how compatibility and conflict-risk terms guide support selection, and how the local response bound relates to the implemented editing procedure. These clarifications distinguish the theoretical properties of the diffusion field from the empirical effects of the full framework.

Together, these additional results provide stronger evidence that graph-guided support selection can improve continual model editing, particularly when related facts interact across successive updates. They also clarify the roles of heat diffusion, historical memory, conflict screening, and the conditions under which their benefits are observed.

We sincerely appreciate the reviewers' time and thoughtful suggestions. We hope these additional experiments and clarifications help address the main concerns, and we would be grateful for any further comments or discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.