acceptodds
Under review as a conference paper at ICLR 2027

Sample-Wise Early Stopping for LLM Unlearning

Abstract

Large language model (LLM) unlearning aims to remove the influence of designated training data while preserving model utility, with applications to privacy protection and regulatory compliance. However, existing methods typically use a fixed training budget, overlooking differences in unlearning difficulty across samples. Consequently, some samples remain insufficiently unlearned, whereas others undergo excessive unlearning, resulting in redundant computation or degraded model utility. To address this challenge, we propose Sample-Wise Early Stopping (SWES), an efficient framework for adaptive sample-wise LLM unlearning. SWES measures each sample’s unlearning progress by comparing its length-normalized negative log-likelihood before and during unlearning. It combines temporal smoothing with threshold-based control to dynamically suspend or resume each sample’s participation in backpropagation, reducing unnecessary updates and focusing subsequent training on samples that have yet to reach the unlearning target. As a plug-and-play framework, SWES can be integrated into a range of gradient-based unlearning methods. Experiments on TOFU and MUSE demonstrate that SWES improves the trade-off between unlearning effectiveness and model utility in multiple settings, while reducing cumulative sample participation in backpropagation by 15.9%–64.6% and achieving training speedups of 1.15×–2.34×.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Machine Unlearning, Early Stopping

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_i8HVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper develops a sample-wise early stopping method, aiming to improve the performance and efficiency of LLM unlearning. The proposed method is a directly plug-in for existing unlearning algorithms. Both performance and efficiency are observed in the experiment results.

Critical strengths
  1. The presentation is good, and the paper is well written. The paper is clear even for people without LLM unlearning background.
  2. In the experiments, multiple LLM unlearning algorithms are included. Overall, the empirical validations are quite strong. The proposed strategy both improves performance and training efficiency.
  3. Studying unlearning/safety training from a data pruning perspective is well motivated.
Critical weaknesses
  1. The proposed method relies on training the entire forget dataset, which is somewhat constrained. This is confined to LLM unlearning's application to removing data memorization (e.g., TOFU and MUSE), where you know the specific texts you want the model to forget. However, if we look at the WMDP benchmark, the forget set is merely a 'proxy' of the harmful knowledge (e.g., how to make Bio weapons) that we want to remove. In this case, I do not think the proposed method can work.
  2. Lack of large-scale and direct comparison with other early stopping methods. While this is validated in Figure 5, this is only a partial evaluation involving a single algorithm.
  3. The method requires selecting a hard threshold for each task and each algorithm, which involves heavy hyperparameter tuning.
  4. While the performance gain is clear, I do not think efficiency improvement for LLM unlearning is a strong motivation. After all, LLM unlearning is actually cheap in compute, where you usually optimize for fewer than 1000 steps.
Questions for authors
  1. In table 1, why not experiment with GD+SWES? Is this because GD and WGA perform similarly? Can SWES also be integrated with representation-level unlearning methods such as RMU? It seems that SWES is naturally more compatible with unlearning algorithms using a reference model?
  2. I do not see why this paper is specific to unlearning. Can it be extended to other tasks, like safety optimization, which also involves constrained optimization.
Reviewer_DJYaOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Sample-Wise Early Stopping (SWES) to reduce redundant computation in gradient-based LLM unlearning. SWES measures each forget sample's progress using the difference in length-normalized negative log-likelihood between the current and original models. A moving-average controller temporarily excludes samples that exceed a suppression threshold from backpropagation and reactivates them if suppression falls below that threshold. The framework is applied to NPO, SimNPO, and WGA, with matched retain batches when needed. On TOFU and MUSE, the authors report 15.9%–64.6% fewer cumulative sample participations in backpropagation and 1.15×–2.34× training speedups, alongside improvements in selected forgetting and utility metrics.

Critical strengths

SWES is straightforward to integrate into existing gradient-based unlearning methods without changing their per-sample objectives. In the reported experiments on TOFU and MUSE with NPO, SimNPO, and WGA, it reduces cumulative backpropagation participation and achieves measurable wall-clock training speedups.

Critical weaknesses
  1. The paper repeatedly motivates SWES by arguing that a uniform training budget may leave some samples insufficiently unlearned while over-updating others. However, heterogeneous sample-wise unlearning progress has already been studied by SatImp (Yang et al., ICML 2025) and BalDRO (Shao et al., WWW 2026). Moreover, Instance-dependent Early Stopping (Yuan et al., ICLR 2025) has already introduced per-sample exclusion from backpropagation and subsequent reactivation. SWES's clearest additional element is its unlearning-specific, reference-relative likelihood stopping criterion, rather than the general stop-and-reactivate mechanism itself. The paper does include an ablation against an IES-inspired second-order-difference criterion (Figure 5), but this is not a full end-to-end comparison of the respective controllers, as reactivation is disabled in that experiment. Nor does the paper directly compare SWES with the closest sample-adaptive unlearning methods, such as SatImp and BalDRO, under matched computational budgets. As a result, it remains unclear whether the proposed stopping signal offers a substantial advantage over established approaches, and the familiar problem motivation alone does not establish methodological novelty.

  2. The threshold τstop\tau_{\mathrm{stop}} is central to SWES because it controls sample participation and computational savings, yet the paper does not provide a principled or transferable procedure for choosing it. Appendix C.1 describes dataset-specific grid searches selected using final forgetting and utility outcomes. Figure 6 does analyze threshold sensitivity, but it does not establish how to choose the threshold without outcome-dependent tuning or whether the same selection rule generalizes across datasets and base methods. More importantly, the abstract and introduction emphasize improvements in the trade-off between unlearning effectiveness and model utility, whereas the experiments do not systematically evaluate or establish a superior forgetting–utility trade-off against the closest sample-adaptive baselines. The reported FQ/MU gains, Figure 6's threshold plots, and Appendix E.3's budget-matched comparison with global early stopping provide some evidence in particular settings, but they do not constitute a comprehensive Pareto-frontier comparison under matched computational budgets. The picture is further complicated by deteriorations in other forgetting measures: on TOFU Forget05, NPO+SWES increases EM from 0.4961 to 0.6957 and ES from 0.0517 to 0.0894 (both lower-is-better), while on MUSE News, WGA+SWES increases VM-Df from 0.0161 to 0.2666. Consequently, the broad trade-off claim highlighted in the abstract and introduction remains insufficiently supported by the experimental analysis.

Questions for authors
  1. Beyond the reference-relative likelihood stopping criterion, what is the substantive methodological difference between SWES and existing sample-adaptive methods? Could the authors extend the IES-inspired ablation into a full controller comparison, including reactivation, and compare against SatImp and BalDRO under matched computational budgets?

  2. How should τstop\tau_{\mathrm{stop}} be selected in practice without tuning against the final forgetting and utility results? Does one threshold-selection rule transfer across forget ratios, datasets, base unlearning methods, and model architectures?

  3. Since the abstract and introduction emphasize an improved forgetting–utility trade-off, could the authors go beyond the existing threshold plots and global early-stopping comparison to report Pareto curves against the closest sample-adaptive baselines under matched budgets? How do they reconcile the improved FQ with worse EM/ES on TOFU and increased VM-Df for WGA on MUSE News?

Reviewer_S6o8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper considers the new angle of unlearning in LLMs. The authors propose a mechanism that dynamically selects samples to be unlearned at each iteration. The experiments show that combining the proposed mechanism with existing unlearning methods can improve the unlearning performance and reduce the running time.

Critical strengths
  • The paper investigates a new angle of unlearning in LLMs
  • The proposed mechanism is simple and can be combined with existing unlearning methods.
Critical weaknesses
  • Limited methodological novelty. The proposed mechanism performs a dynamic sample selection, which is a well-established technique in machine learning [1]. Using the reference-relative likelihood suppression as the selection criterion is also not novel [2]. The main contribution of the paper is to combine these two techniques and apply them to unlearning in LLMs.

  • Insufficient evaluation. Experiments are conducted on only a single LLM and single run.

  • Limited analysis. The paper can perform more analysis to understand the proposed mechanism.

    • Is there any correlation between the stopping scores of samples and their actual unlearning performance; any sample is considered as stopped but remaining extractable.
    • Why a shared threshold is used across heterogeneous samples? Sample difficulty can vary significantly, and a shared threshold may not be optimal.
    • Missing analyzing the sensitivity of the smoothing window size
    • The run time breakdown can be beneficial. The proposed method requires additional inference and loading of the reference model, which can be expensive. Analyzing both running time and memory overhead can provide a more comprehensive picture. Analyzing the convergence of the methods and the required number of iterations can also be beneficial in understanding the efficiency.

[1] https://arxiv.org/abs/2206.07137 [2] https://arxiv.org/abs/2404.07965

Questions for authors
  • How does the baselines are tuned? Why the baselines never match the FQ of SWES (intuitively, if we run the baselines more, we should match the FQ of SWES)? How does the trade-off between unlearning performance and model utility of the baselines compare with that of SWES?

  • Should this method be called "dynamic sample selection" instead of "early stopping", which is a bit misleading. Early stopping is a well-established technique in machine learning, which is different from the proposed method.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.