acceptodds
Under review as a conference paper at ICLR 2027

Fast and Memory-Efficient Lifelong Model Editing with Certified Evaluation Reuse

12 views

Abstract

Model editing writes new knowledge into a frozen large language model without retraining. In lifelong editing, edits arrive continually and must be re-verified against the accumulated history. A single edit in a memory-based editor changes one active value row, yet the standard execution path recomputes edit-invariant prefixes and retrieval at every optimization step, allocates training state for the whole table, and re-tests the full history at every checkpoint. We propose eHoReN, which removes this mismatch in editors that retrieve before they inject and write one row, with a single principle: the execution scope of every computation and every piece of storage should equal its dependency scope on mutable state. Each technique carries an equivalence proposition fixing when the original behavior is preserved. On the editor that holds the quality frontier at scale, eHoReN edits up to 3.5x faster, with outputs identical to the baseline on Llama-3.1-8B and quality within seed variability on Qwen2.5-7B. Its training state shrinks from 4.58 GiB to 48 KiB and stops growing with the memory. The same dependency records make reuse of historical evaluation results checkable, and a cost model predicts when checking pays off. Certified reuse speeds up evaluation by 1.89x per checkpoint without lag, and periodic re-certification reaches 1.6-2.4x over exact checking with exact agreement at every anchor. Transferred to a second editor with different retrieval and injection mechanisms, the editing speedup reproduces and the evaluation gain follows the benefit condition's prediction. Code is available at: https://anonymous.4open.science/r/eHoReN/

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
model editing, knowledge editing, lifelong model editing, efficiency, memory-based editing, evaluation reuse, incremental computation, large language models

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 2 reviews, above 27% of the others.
Reviewer_1CiNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes eHoReN, based on a simple observation: each edit only modifies one row in the external memory, but existing implementations repeatedly compute many things that are not affected by this edit. The authors divide these redundancies into three parts: editing computation, training states, and historical evaluation, and optimize them separately. The goal is to make continuously updating knowledge in LLMs and checking whether previous knowledge is affected faster and more memory-efficient.

Critical strengths
  1. Clear motivation and simple idea: The motivation is clear, and the core idea is simple and intuitive.
  2. Solid engineering: The method is solid and reasonable from an engineering perspective. The authors identify several engineering issues in methods like HoReN and fix them.
  3. Efficiency: The method improves efficiency with almost no performance loss, making model editing much more scalable.
Critical weaknesses
  1. Novelty: Although the engineering is solid, most of the improvements are engineering-level fixes, and the novelty is limited.
  2. Generality: The efficiency improvements mainly apply to methods like HoReN and are not general to all editing methods. Although the experiments on GRACE provide some support for generality, not all editing methods have these issues.
  3. Experimental results: I noticed that the main experiments (Table 1) only include Llama-3.1 8B / ZsRE and Qwen2.5 7B / WikiBigEdit. Also, the results on ZsRE and WikiBigEdit seem quite different. AlphaEdit performs much better than other methods on WikiBigEdit, but the paper does not explain this difference well.
  4. Missing evaluations: The paper does not report the Persona and multi-hop reasoning scores provided by the original WikiBigEdit benchmark.
Questions for authors
  1. Could the authors provide more detailed results for Llama-3.1 8B / WikiBigEdit and Qwen2.5 7B / ZsRE?
  2. Could the authors explain why the results on ZsRE and WikiBigEdit are so different?
  3. Could the authors provide the Persona and multi-hop reasoning scores on WikiBigEdit?
Reviewer_qzbQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes eHoReN, an execution optimization framework for lifelong model editing that aims to reduce editing costs without changing the original editing algorithm. The method follows the principle that the scope of computation and storage should match their dependency on mutable state. Specifically, it introduces frozen-prefix computation reuse, within-edit route caching, supervised-position projection, and active-row training to accelerate individual edits and reduce training-state memory. It also proposes route-aware evaluation reuse, which checks whether the effective injection sequence of a historical query has changed before deciding whether to regenerate its output. The paper provides equivalence propositions for the proposed optimizations and evaluates them on two backbone models, two datasets, multiple editing baselines, and long editing settings. The results demonstrate editing speedups and reduced training-state memory, while the evaluation reuse strategy can reduce repeated evaluation costs under the stated correctness conditions.

Critical strengths
  1. Clear motivation: The paper addresses redundant computation and repeated evaluation in lifelong model editing, which are important efficiency bottlenecks.
  2. Promising efficiency gains: The paired experiments report up to 3.51× editing speedup, together with substantial reductions in optimizer-related training-state memory.
  3. Long-horizon evaluation: Experiments extending to 10K edits provide useful evidence of the method's scalability.
Critical weaknesses
  1. Limited experimental coverage: The evaluation uses only two backbone models and two datasets. Although several editing baselines are included, the proposed optimizations are mainly validated on HoReN and GRACE, leaving their applicability to other editing mechanisms insufficiently explored. e.g RL related knowledge editing
  2. Limited hyperparameter analysis: Although component-level ablations are provided, the sensitivity of editing quality, runtime, and memory usage to key hyperparameters is not systematically investigated.
  3. Restricted correctness assumptions: The theoretical guarantees rely on specific execution conditions, including fixed configurations and deterministic behavior. The applicability of these guarantees to more general deployment settings remains limited.
  4. Outdated related work: The Related Work section does not adequately cover recent advances in knowledge editing, particularly methods published in the past two years.
Questions for authors
  1. Improve the method illustration: Figure 3 could more clearly distinguish within-edit computation reuse from cross-checkpoint evaluation reuse, including their cache contents and recomputation conditions.
  2. Discuss caching trade-offs: A more consolidated analysis of cache memory consumption versus evaluation-time savings would clarify the practical applicability of route-aware reuse.
  3. Clarify comparison protocols: The paper should consistently distinguish paired speedup comparisons within the same editor from timing comparisons across different editing methods with different experimental budgets.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.