acceptodds
Under review as a conference paper at ICLR 2027

Forget, But Not Too Much: Self-Limiting LLM Unlearning

Abstract

Machine unlearning aims to remove selected knowledge from a trained model without retraining from scratch. Most existing LLM unlearning methods reduce the probability of the target answers. This does not match how TOFU evaluates forgetting. TOFU compares the unlearned model with a model that never saw the forget set, so pushing the target answers far below that reference can reduce Forget Quality just as retaining them can. We show that common one-sided objectives have no stationary point at this target. They can therefore keep updating after the model has reached its best agreement with the reference. We introduce NGREF, a self-limiting objective that matches the model to its frozen pretrained checkpoint on the forget set while retaining performance on the remaining data. The pretrained checkpoint becomes a fixed point of the forget objective. On the full published TOFU comparison grid with two models, three forget splits and four LoRA ranks, NGREF exceeds the best Forget Quality reported by nine published methods in all 24 settings. A controlled objective swap raises Forget Quality from approximately zero to 0.585 while leaving model utility similar. In a training sweep, a strong one-sided baseline reaches 0.766 and then falls to 0.0013 even though its forget loss continues to improve. Reference matching instead reaches 0.99 and remains close to that value over the measured range. These results show that the endpoint of unlearning matters as much as the strength of forgetting. Code: https://anonymous.4open.science/r/ngref-unlearn-4F8E/README.md

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
machine unlearning, large language models, knowledge removal, knowledge unlearning, parameter-efficient fine-tuning

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_H5ogOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates a mismatch between common LLM unlearning objectives and reference-based evaluation. One-sided objectives continue suppressing forget-answer probabilities even after the model approaches the behavior of a model that never learned the forgotten data. The authors propose NGREF, which minimizes token-level KL divergence from a frozen pretrained checkpoint on forget examples while maintaining a cross-entropy objective on retain examples. The paper provides a distribution-space analysis and evaluates the method across 24 TOFU configurations, supplemented by training trajectories, controlled objective comparisons, preconditioner ablations, and a secondary extraction experiment.

Critical strengths
  1. The distinction between suppressing an answer and recovering the behavior of an ignorant reference is well motivated.
  2. The evaluation covers two models, three forget splits, and four LoRA ranks.
Critical weaknesses
  1. The figures use inconsistent formatting and have poor readability. Please standardize their presentation and improve the legibility of labels and annotations.
  2. NGREF matches the pretrained checkpoint, whereas TOFU evaluates against a retain-only model. The suitability of this surrogate needs stronger empirical justification.
  3. The baseline performs well before over-suppression. Comparisons under a shared stopping protocol would better establish NGREF’s practical advantage.
  4. The objective uses standard KL distillation and retain-set training. Its distinction from existing reference-based unlearning methods should be clarified.
Questions for authors

see weakness

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

This paper raises the attention that continually reducing the probability of the target answers might lead to overshoot. Therefore, it proposes NGREF, a self-limiting objective that matches the model to its frozen pre-trained checkpoint on the forget set while retaining performance on the remaining data. Experiments on TOFU demonstrate its effectiveness.

Critical strengths
  1. The over-forgetting observation is reasonable and worths attention.
  2. NGREF is simple to implement and gives competitive empirical results under the TOFU benchmark.
Critical weaknesses
  1. To my knowledge, over-forgetting or over-unlearning has already been explicitly recognized in prior work [1,2]. Thereofre, the insight that unlearning requires a bounded or target-seeking objective, rather than indefinitely reducing forget-answer probabilities, has been pointed out before. From this perspective, the main distinction of NGREF appears to be its specific choice of the pre-task checkpoint as the reference distribution, with a limited conceptual novelty.
  2. More importantly, the paper does not adequately justify why the checkpoint before task fine-tuning should approximate the predictions of a model retrained without the forget data. Proposition 2 only shows that KL is minimized at the selected teacher, which does not validate this teacher as the correct unlearning target. Overall, I think the authors could/should provide more substantial theoretical and empirical evidences for this surrogate. If this is only limited to TOFU benchmark, I am afraid this surrogate is limited for practical unlearning we have little knowledge about the property of forget set and the start point of model to be unlearned.

[1] Towards Unbounded Machine Unlearning, 2023, NeurIPS.

[2] UNDIAL: Self-Distillation with Adjusted Logits for Robust Unlearning in Large Language Models, 2025, NAACL.

Questions for authors
  1. Could you provide some verifiable assumptions that would make the pre-trained checkpoint a consistent proxy for the retain-only model?
Reviewer_st16OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper identifies a mismatch between common unlearning objectives and the counterfactual target used by TOFU. The authors propose NGREF, which uses KL divergence to match the unlearned model's output distribution to a frozen pretrained checkpoint, combined with a retain loss. Experiments on TOFU show improvements over published baselines, with additional analyses of training dynamics and a secondary extraction experiment.

Critical strengths
  1. The paper identifies a limitation of several common unlearning objectives under reference-based evaluation.
  2. The proposed objective is simple and computationally practical, and the experiments show improvements across different TOFU configurations.
Critical weaknesses
  1. My main concern is that the paper's proposed approach seems closely tied to one specific benchmark (TOFU). The main evaluation remains heavily focused on TOFU, and the proposed objective seems largely motivated by the specific behavior rewarded by TOFU's Forget Quality metric. I'm not sure how convincing this motivation is in the first place, and how well these findings generalize beyond this particular benchmark.

  2. Another concern is that the proposed method uses the pretrained checkpoint as a reference for the desired unlearned behavior. This assumption may hold for TOFU, where the forget knowledge is introduced during fine-tuning, but may not hold in more general unlearning settings, e.g. when the forget information already exists in the pretrained model.

  3. I also find the methodological contribution a bit incremental. The proposed objective essentially combines KL-based reference matching with a retain loss. Related ideas (knowledge distillation and reference-based unlearning) have been explored a lot in previous work [1, 2, 3]. I would appreciate a discussion of how this method differs from these existing methods and what its main novelty is. Moreover, the main results are compared against previously published baseline numbers, rather than re-evaluating those baselines under the same experimental setup, which is a bit questionable.

[1] Towards unbounded machine unlearning. Kurmanji et al. NeurIPS 2023. [2] KGA: A General Machine Unlearning Framework Based on Knowledge Gap Alignment. Wang et al. ACL 2023. [3] Reference-Guided Machine Unlearning. Mirlach et al. 2026

Questions for authors

See "Critical Weaknesses".

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.