acceptodds
Under review as a conference paper at ICLR 2027

BEYOND FIRST CROSSINGS: PRESERVATION IN FACTUAL REWRITING AT MATCHED MEAN OPTIMIZATION-PROMPT LOSS

Abstract

We study which preservation differences in factual rewriting persist after match- ing the mean negative log-likelihood (NLL) achieved on optimization prompts. Starting from the same complete intermediate state, we apply suffix programs with different learning rates and compare discrete first crossings with directly evaluated parameter states within the final update. An initial exposure-matched load–cost reversal is absent under matched base-training steps, motivating these within-state comparisons. In five new controlled initialization–mapping pairs, subsequent mean-loss matching retains high-load old net-cost reductions of ap- proximately 1.15 and 1.07 nats for slow relative to fast suffixes across two pre- fixes. Here, old net cost is the change in mean old-target NLL from the base model. On Qwen3-8B-Base/CounterFact, mean coarse-step contrast magnitudes shrink from roughly 0.03–0.06 nats to approximately 0.0063 nats or less, whereas fine-step comparisons retain much smaller differences of approximately 0.002– 0.004 nats after extending eligible trajectories and narrowing the matching band. In the fine-step narrow-band comparison, slow suffixes have lower mean old net cost but higher mean held-out test NLL on complete-expression batches, despite equal test-accuracy counts in every paired batch. Thus, the large controlled dif- ferences survive mean-loss matching, whereas the coarse natural-fact contrasts largely do not. Per-edit losses and independent-expression effectiveness can still differ at a matched mean optimization loss.

Primary area:
interpretability and explainable AI
Keywords:
Knowledge Editing, Factual Rewriting, Knowledge Preservation, Model Evaluation, Optimization Dynamics, Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_tEywOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether preservation differences between slow- and fast-learning-rate factual rewriting programs persist after matching the mean optimization-prompt NLL of their terminal states. The two suffix programs start from the same saved model, optimizer, and sampling state. The authors compare the checkpoints at which each program first crosses an editing-loss threshold with fractional states selected within the final update. In a controlled Transformer, large slow–fast preservation differences remain after matching. On Qwen3-8B-Base and CounterFact, coarse-step differences largely disappear, while fine-step experiments retain much smaller differences. The held-out analysis further shows that lower old-fact cost for the slow suffix can coexist with higher test NLL and equal test accuracy.

Critical strengths
  1. The paper identifies an important evaluation issue in factual rewriting: a shared first-crossing threshold can select states with substantially different achieved editing losses. Comparing first-crossing states with directly evaluated mean-loss-matched states provides a useful diagnostic for separating endpoint overshoot from preservation differences.

  2. The experiments are detailed and extensive with respect to this evaluation question. The authors examine controlled and natural-fact settings, coarse- and fine-step programs, matching failures, per-edit differences, and held-out evaluation; in particular, the attenuation of the coarse-step Qwen differences and the held-out test-NLL trade-off make the conclusions more transparent and informative.

Critical weaknesses
  1. The controlled study defines factual load using 200 or 3,000 base-training associations, but does not report full-set pre-edit accuracy or NLL, including performance on the 20 edited subjects and the additional 2,800 high-load associations. These additional facts are also excluded from the main preservation pools. Thus, the analysis does not verify whether nominal factual load reflects acquired knowledge or whether the reported preservation results characterize forgetting across the high-load setting.

  2. The matched-state experiments focus on mean optimization NLL near a single target of 0.01, without establishing whether the conclusions hold at other editing-strength levels. Moreover, matching the mean does not ensure equivalent per-edit or held-out effectiveness, as the authors also acknowledge. M1M_1 only narrows the tolerance around the same target and therefore does not address either limitation.

Questions for authors
  1. Can the authors report pre-edit accuracy and NLL over all 200 or 3,000 base associations, separately for the shared facts, the 20 edited subjects, and the additional 2,800 high-load associations? How many of these associations are reliably mastered before editing?

  2. What motivates the target NLL of 0.01 and the matching bands centered at 0.00995? Do the slow–fast preservation contrasts persist when terminal states are matched at other editing-strength levels?

  3. How do the conclusions change when effectiveness is compared at the per-edit level or on held-out expressions, rather than only through batch-mean optimization NLL?

  4. Do the main attenuation results persist for another model, a larger edit batch, or a less restricted parameter-update setting?

Reviewer_d7rfOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates how different evaluation endpoints affect conclusions about knowledge preservation in factual rewriting. The authors find that comparing editing methods at their first crossing of a common loss threshold can be misleading, as different optimization schedules may achieve substantially different final editing losses. To take this issue, they propose a controlled evaluation protocol that branches from the same intermediate model and optimizer state.

Critical strengths
  1. The paper highlights a subtle but important issue in knowledge editing evaluation: reaching the same loss threshold does not guarantee comparable editing outcomes.

  2. The same-state suffix intervention and matched-loss evaluation help control confounding factors and enable more meaningful comparisons.

  3. Experiments on both controlled Transformers and Qwen3-8B, together with sensitivity analyses, provide useful insights into preservation differences across settings.

Critical weaknesses
  1. The experiments focus on gradient-based editing with different learning rates, without comparisons to representative knowledge editing methods such as ROME, MEMIT, or AlphaEdit. This limits the generalizability of the findings.

  2. The evaluation relies heavily on NLL. Although accuracy is reported, standard knowledge editing metrics such as edit success, generalization, and locality are not systematically evaluated.

  3. The preservation differences on Qwen3-8B become very small after loss matching, while held-out test accuracy remains unchanged. The practical significance of these improvements is unclear.

Questions for authors
  1. Can the authors validate whether the conclusions generalize to other knowledge editing methods, such as ROME or MEMIT?

  2. Can the authors provide a more comprehensive evaluation using standard knowledge editing metrics, particularly editing success rate and locality?

  3. Since matching mean NLL does not guarantee matching per-edit losses or optimization steps, how much of the remaining preservation difference can be attributed to these factors?

Reviewer_Fv3ROpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies which preservation differences in factual rewriting persist after matching the mean negative log-likelihood (NLL) achieved on optimization prompts. It observes that: 1) the large high-load preservation differences persist after matching mean optimization-prompt NLL. 2) mean-loss matching sharply reduces coarse-step preservation differences, while fine-step comparisons retain much smaller residuals after extending eligible trajectories and narrowing the matching band. 3) lower mean old net cost accompanies higher mean held-out test NLL, despite closely matched mean optimization-prompt NLL and equal test-accuracy counts in every paired batch.

Critical strengths
  1. Understanding how different learning strategies affect factual acquisition and preservation could provide useful insights in model continual learning.
Critical weaknesses
  1. Primarily empirical observations with limited explanatory insight. My main concern is that the paper mostly reports empirical observations, but provides limited theoretical or mechanistic explanation for why these differences arise.

  2. Several experimental design choices are insufficiently motivated. A number of experimental choices appear somewhat arbitrary or are not sufficiently justified. For example, it is unclear why the analysis focuses on five pairs, why the particular two learning-rate categories are chosen, and how other experimental hyperparameters are determined.

  3. Limited evidence for generality. The experiments appear to be conducted only on Qwen3-8B-Base and CounterFact. This makes it difficult to determine whether the reported phenomena reflect a general property of factual rewriting and optimization or are specific to this particular model-dataset combination. The claims would be significantly strengthened by experiments across different model families, model scales, datasets, or factual editing settings.

  4. The practical implications for learning-strategy design remain unclear. Although the paper observes differences among learning strategies after controlling for NLL, it is unclear how these findings translate into actionable guidance. Can the identified metrics predict which learning strategy will preserve existing knowledge better? Can the observations be used to select learning rates, optimization procedures, or stopping criteria?

Questions for authors
  1. What is the authors' hypothesis for why substantial preservation differences remain even after matching optimization-prompt NLL?

  2. What motivates the choice of five matched pairs, the two learning-rate categories, and the specific hyperparameters? How sensitive are the main conclusions to these choices?

  3. Do the same observations hold for other model families, model sizes, and factual rewriting datasets beyond Qwen3-8B-Base and CounterFact? Even a smaller-scale validation on an additional model and dataset would substantially strengthen the generality claim.

  4. How should these observations inform the design or selection of learning strategies?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.