acceptodds
Under review as a conference paper at ICLR 2027

RECON: Retrieval-Conditioned Norm Calibration for Lifelong Model Editing

Abstract

Lifelong model editing requires a language model to absorb a growing sequence of factual updates while preserving prior knowledge and unrelated behavior. Residual-memory editing enables such scalable updates by storing each edit as a sparse residual. At scale, however, the hash-based readout of existing residual-memory editors induces two failure modes: routing fragility, because hash lookup is not invariant under paraphrase, and norm inflation, because collisions on shared columns become inevitable over long edit sequences. We present RECON, which repairs the readout with two complementary training-free mechanisms. First, semantic routing–support decoupling replaces rigid hash lookup with a gated dense semantic key while executing the sparse support verbatim, restoring robust retrieval across paraphrased queries. Second, retrieval-conditioned norm calibration rescales each retrieved dimension by a power of its relative column norm, directly counteracting collision-induced magnitude distortion to sustain locality under sequential edits. Across temporal, hallucination-correction, and sequential editing benchmarks, RECON nearly doubles out-of-distribution generalization over the hash-based readout at , preserves locality across all evaluated horizons, and retains its advantages as edit sequences grow.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Lifelong Model Editing, Residual Memory, Semantic Routing, Norm Calibration

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_4fVyOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper studies performance degradation in lifelong knowledge editing for LLMs. The authors identify inference-time memory retrieval and execution as important bottlenecks in existing residual-memory editing methods. To address these issues, they propose RECON, which replaces hash-based retrieval with semantic routing and introduces Retrieval-Conditioned Norm Calibration to mitigate interference caused by write collisions in shared residual parameters. Experiments demonstrate improvements in OOD generalization and long-horizon sequential editing.

Critical strengths
  1. The paper offers a useful perspective on the performance bottlenecks of residual-memory editing by focusing on memory readout. This provides insights into the limitations of existing methods and potential directions for improvement. RECON is also lightweight, requiring no changes to the original write process or additional training of the retrieval module.

  2. The experiments cover multiple models, benchmarks, and editing horizons, with promising results in OOD generalization and long-horizon sequential editing.

Critical weaknesses
  1. Although the ablation studies show the effectiveness of semantic routing, the paper lacks a systematic analysis of when and why the original hash-based routing fails, and to what extent these failures contribute to editing performance degradation.

  2. The authors observe a correlation between write collisions and residual column norms. However, the current theoretical and empirical evidence does not sufficiently establish whether norm inflation is a key cause of editing performance degradation, nor does it directly demonstrate how the proposed calibration mitigates interference among different edits.

  3. Semantic routing requires storing semantic representations for all edits and performing similarity-based retrieval. At larger editing scales, its retrieval latency is substantially higher than that of hash-based routing. Although the authors suggest approximate nearest-neighbor search as a potential acceleration strategy, they do not evaluate it in practice, leaving the trade-off between computational overhead and performance gains unclear.

Questions for authors
  1. I would like to see a more detailed analysis of the relationship between retrieval errors and editing failures, particularly for the original hash-based routing. Comparisons across different semantic representations and similarity metrics would also help establish the necessity of the proposed routing design. Similarly, further analysis of whether norm inflation is a key factor in editing degradation would strengthen the motivation for Norm Calibration.

  2. The largest gains from Semantic Routing are observed on Temporal, which contains at most 75 sequential edits. In contrast, its contribution appears relatively limited in the 15,000-edit setting on ZsRE. Does this suggest that Semantic Routing is primarily beneficial under distribution shift rather than at longer editing horizons? How would the method perform when long editing sequences and OOD queries are considered together?

Reviewer_9iTyOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents RECON, a readout-focused approach to lifelong model editing built upon sparse residual memory. The authors identify two limitations of existing hash-based retrieval: sensitivity to query reformulation and magnitude distortion caused by overlapping residual updates. RECON addresses these issues through semantic routing with support-preserving execution and retrieval-conditioned norm calibration, without introducing additional training for the readout mechanism.

Experiments on ZsRE, Temporal, and SelfCheckGPT demonstrate improvements over MEMOIR, particularly under distribution shift and long editing sequences. Notably, RECON improves Temporal generalization from 0.37 to 0.72 at (T=10) and maintains strong locality at 15,000 edits.

Critical strengths
  1. The paper identifies two key limitations of residual-memory editing, namely fragile hash-based retrieval and collision-induced norm inflation. The analysis is clear and supported by empirical evidence, providing a convincing motivation for the proposed method.

  2. RECON offers a simple yet effective solution by combining semantic routing with norm calibration. Both components are training-free and preserve the original sparse writing mechanism, making the approach practical and easy to integrate into existing frameworks.

  3. The experimental results are promising, particularly the substantial improvement in out-of-distribution generalization and the stable performance over 15,000 edits. The ablation studies also help validate the effectiveness of the proposed components.

Critical weaknesses
  1. The generalizability of semantic routing could be better established. The strongest improvements are observed on Temporal, which contains only 75 edits, while the performance gains become smaller at longer editing horizons. Additional evaluations under more diverse distribution shifts would strengthen the robustness claims.

  2. The contribution of norm calibration is relatively modest. Although the ablation results support its effectiveness, the average improvement is only around 0.55 percentage points at T=1000 on ZsRE. Comparisons with simpler normalization or scaling strategies would help clarify its added value.

  3. The scalability benefits come with additional retrieval costs. At 15,000 edits, semantic routing requires substantially more time than hash-based retrieval and introduces extra memory overhead. Further evaluation of the accuracy–efficiency trade-off would make the practical advantages more convincing.

Questions for authors
  1. The semantic gate threshold varies across datasets and editing horizons. How sensitive is RECON to these choices, and would a fixed threshold maintain similar performance under unseen distribution shifts?

  2. Norm calibration provides consistent but relatively modest improvements. Have the authors considered comparing it with simpler alternatives, such as norm clipping or uniform residual scaling, to better demonstrate its specific advantages?

  3. Semantic routing introduces additional memory and inference overhead as the number of edits increases. Could the authors provide experimental results using approximate nearest-neighbor search or key compression to clarify the practical scalability of RECON?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.