acceptodds
Under review as a conference paper at ICLR 2027

Explicit Representation Alignment via Subspace Elimination for LLM Unlearning

Abstract

Large Language Models (LLMs) may retain targeted knowledge, including sensitive or hazardous information, even when such knowledge is undesirable in deployment. Machine unlearning aims to suppress the influence of selected data while preserving useful model behavior. We present ERASER, a representation-guided unlearning method that combines a data-derived subspace target with a ranking-based Disentangle-Head. ERASER constructs a fixed aggregate forgetting projector offline from predefined forget groups and uses it to form a sample-specific residual target during training, while ranking encourages higher learned scores for retain samples than for forget samples. We evaluate ERASER on WMDP, TOFU, MUSE, and Fictional Knowledge, observing favorable forgetting-retention trade-offs relative to the evaluated baselines. On Fictional Knowledge, it reduces recovery on exact, paraphrased, and compositional probes. Under the tested protocols, ERASER also yields lower jailbreak attack success rates than the pre-unlearning model and RMU. We additionally evaluate forget-set membership inference and sensitivity to limited-data relearning.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LLM Unlearning, Alignment Learning, Pairwise Ranking

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_5tvgOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ERASER, a representation-guided method for LLM unlearning. It constructs a fixed forgetting subspace from frozen-model activations and penalizes the components of forget representations that lie in this subspace, while a ranking-based Disentangle-Head encourages retain samples to receive higher scores than forget samples. Experiments on WMDP, TOFU, MUSE, and Fictional Knowledge report competitive forgetting–retention trade-offs, along with additional evaluations of jailbreak attacks, membership inference, and limited-data relearning.

Critical strengths
  • ERASER replaces RMU’s random target with a data-derived forgetting subspace, providing a more interpretable representation-level objective without requiring a manually selected target scale.

  • The evaluation is broad, covering four unlearning benchmarks, multiple model families, component ablations, abstraction-level probes, jailbreak attacks, membership inference, and limited-data relearning.

Critical weaknesses

i) ERASER constructs its forgetting subspace from the leading directions of uncentered forget-set activations, without contrasting them with retain-set activations. These directions may therefore capture common high-variance language features rather than knowledge specific to the forget set. Unlike LEACE, which defines concept erasure through the inability of linear classifiers to recover the target concept while minimizing changes to the remaining representation, ERASER does not provide an operational test showing that its projected directions encode the knowledge to be removed. Prior work has also shown that projection-based removal may unintentionally affect information unrelated to the target [a]. This concern is reflected in Appendix B.6, where unlearning Economics and Law reduces accuracy on the adjacent Econometrics and Jurisprudence tasks by 13.16 and 23.15 percentage points, respectively.

The ablation results further suggest that most of the improvement comes from the Disentangle-Head rather than the proposed projector. In Table 12, adding the same head to RMU already produces results close to the full ERASER method: ERASER improves Bio and Cyber accuracy by only 0.71 and 1.31 percentage points over RMU with the head, while slightly reducing MMLU. Moreover, the ranking head can learn to separate forget and retain samples through its own parameters without making the underlying representation space geometrically disentangled. A stronger comparison against RMU with the same head, together with direct measurements of target-information recoverability and forget-retain subspace overlap, is needed to establish the contribution of subspace elimination.

ii) The membership-inference experiment compares forget samples with controls drawn from different corpora, so it cannot separate membership information from differences in data source or content. The paper acknowledges this limitation, but the resulting scores provide little evidence about privacy after unlearning. The jailbreak results are also difficult to interpret because the qualitative examples show that ERASER often produces fragmented or malformed responses. Its lower GCG attack success rate may therefore reflect degraded output quality rather than removal of the target knowledge. This distinction matters because output-level suppression can coexist with recoverable information in the model’s internal representations [c]. Under the embedding-space attack, the cumulative attack success rate still reaches 78%, further limiting the robustness claim.

The relearning evaluation is similarly incomplete. It fine-tunes only ERASER on 10, 30, or 50 examples from the original forget set and does not compare its recovery with RMU or other unlearning methods. It also omits retain-data and unrelated-data controls. Fine-tuning on retain data would show whether ordinary continued training can reactivate the target knowledge, while unrelated data would help determine whether any recovery reflects residual knowledge or merely a general effect of additional fine-tuning. Prior work has shown that apparently forgotten knowledge can reappear after benign relearning [b], making these controls and baseline comparisons important before claiming stronger resistance to recovery.

[a] Better Hit the Nail on the Head than Beat around the Bush: Removing Protected Attributes with a Single Projection. EMNLP 2022.

[b] Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning. ICLR 2025.

[c] An Illusion of Unlearning? Assessing Machine Unlearning Through Internal Representations. AISTATS 2026.

Questions for authors
  1. Can the authors directly test whether the projected subspace contains forget-specific knowledge rather than common high-variance features, for example through linear probes, forget-retain subspace overlap, or comparison with a retain-conditioned concept-erasure method?

  2. In Table 12, RMU with the Disentangle-Head performs very similarly to the full ERASER method. Are the remaining differences statistically significant across random seeds, and what evidence shows that the subspace projector contributes meaningfully beyond the ranking loss?

  3. How does ERASER compare with RMU and other baselines under the same relearning protocol? Can the authors also include retain-data and unrelated-data controls to determine whether recovery is specific to residual target knowledge or a general consequence of additional fine-tuning?

Reviewer_L1eNOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents ERASER, a representation-based, robust LLM unlearning method that combines a data-derived subspace target with a ranking-based Disentangle-Head. ERASER constructs a fixed aggregate forgetting projector offline from predefined forget groups and uses it to form a sample-specific residual target during training, while ranking encourages higher learned scores for retain samples than for forget samples. ERASER is evaluated on WMDP, TOFU, MUSE, and Fictional Knowledge. The authors claim that achieving favorable forgetting-retention trade-offs compared with the evaluated baselines also leads to improved robustness against knowledge recovery attacks.

Critical strengths

None

Critical weaknesses

1. The motivation is not convincing.

The Introduction motivates ERASER as follows (lines 37–41):

"In RMU-style formulations, forget representations are driven toward a prescribed target vector while retain representations are regularized toward their frozen-model counterparts. This formulation uses random target directions with separately chosen scales and optimizes forget and retain representations without explicitly contrasting them."

I do not find this argument convincing. RMU forget loss pushes forget-samples' representations toward a scaled random vector cuc\mathbf{u}, and RMU retain loss keeps retain-samples' representations close to those of the frozen model. Together, the two losses already push forget representations away from retain representations; RMU contrasts them implicitly by design.

The paper goes on to say:

"In practice, target construction and the relative weighting of these objectives can affect the resulting forgetting-retention trade-off."

The same is true of ERASER. RMU requires tuning two hyperparameters (the scaling coefficient cc and the retain weight α\alpha). ERASER requires tuning three hyperparameters, i.e., α\alpha, β\beta, and the subspace dimension kk. ERASER therefore does not remove the tuning burden it identifies as RMU's weakness.

The improvement over RMU is also not clearly significant. In Table 2 (WMDP), ERASER's MMLU equals RMU's (56.6), and its WMDP accuracy is lower by about 3%. I am not clear why an additional loss (the Disentangle-Head) is needed. Why should a target subspace be recomputed from activations for every new forget request, when RMU uses a random vector and costs nothing and already works well? What concrete failure of RMU does ERASER fix that cannot be fixed by tuning cc?

2. Inconsistent use of comparison baselines in experiments. I noticed inconsistent baseline selections across benchmarks. Specifically, Table 2 (left up) compares against five baselines (SCRUB, SSD, RMU, NPO, SimNPO) on WMDP, whereas Table 2 (right up) (TOFU) compares against a different set of methods (GradDiff, NPO, RMU, SimNPO, DPO). The authors should clarify why different baselines are used across benchmarks and whether all methods could be evaluated under a unified setting for a fairer comparison.

3. The comparison may be unfair.

RMU and ERASER both optimize a forget loss with a retain loss (ERASER even has a ranking loss). It is unclear whether NPO and SimNPO were run with a retain loss (e.g., NPO+KL) or with the forget loss only. The authors should clarify the exact objective and hyperparameter search used for each baseline.

4. Is the paper about robustness or about the forget–retain trade-off?

There is tension in the paper. Is it proposing a new robust unlearning method, or an incremental method for improving the forget-retain trade-off? The name stands for "Explicit Representation Alignment via Subspace Elimination for Robust LLM Unlearning" (lines 42-43), but the title is "Explicit Representation Alignment via Subspace Elimination for LLM Unlearning".

If robustness is the central claim, then its evaluation is quite narrow: The relearning experiment (Table 5) reports only ERASER's performance. Several commonly used recovery attacks are also not included [1].

[1] Łucki, Jakub, et al. "An adversarial perspective on machine unlearning for ai safety." arXiv preprint arXiv:2409.18025 (2024).

Questions for authors

Please see the Weaknesses. I think the paper needs a major revision.

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

The paper focuses on representation-guided LLM unlearning. The challenge is that RMU-style methods use a random target vector with separately chosen scales and do not directly compare forget and retain samples during training. To address this, the authors propose ERASER, which constructs a fixed forgetting subspace offline from the frozen model's activations for forget groups and penalizes the projection of forget representations onto this subspace. It further uses a ranking-based Disentangle-Head that assigns higher scores to retain samples than to forget samples. Experiments on WMDP, TOFU, MUSE, and Fictional Knowledge show better forgetting–retention trade-offs than the evaluated baselines and lower jailbreak success rates than the base model and RMU.

Critical strengths

The motivation is clear. Table 1 shows that RMU's forget accuracy is highly sensitive to c, and a data-derived target is a good way to remove this hyperparameter sensitivity.

The evaluation is broad and comprehensive, covering four benchmarks, several model families, jailbreak attacks, membership inference, and relearning.

Critical weaknesses

The contribution of the subspace target appears limited. Table 12 shows that ERASER without the head achieves 64.37% on WMDP-Bio, compared with 63.7% for the base model, indicating essentially unchanged performance. RMU+Head already achieves 28.44% on WMDP-Bio, and ERASER improves on this by only 0.71 percentage points. Thus, most of the gain appears to come from the head rather than the subspace target. Moreover, removing a linear subspace is not a new idea, and pairwise ranking is a common loss formulation. Adding the head to RMU yields similar results, suggesting that it may function as a general-purpose plug-in independent of the specific subspace design. If the two components are indeed interdependent, the authors should clarify their relationship. As presented, the emphasis on "Subspace Elimination" in the title and the claim that the two components are "coupled" are not sufficiently justified. The paper should clarify its technical contributions.

Several important baselines are missing [1, 2, 3]. The evaluated baselines are not representation-based and are relatively simple and old. The paper should clarify what is new beyond using projection as a training loss and whether the proposed subspace design outperforms existing methods, given that this is also presented as a contribution.

[1] Hu, Jinwei, et al. "Falcon: Fine-grained activation manipulation by contrastive orthogonal unalignment for large language model." Advances in Neural Information Processing Systems 38 (2026): 102210-102232.

[2] Wang, Xu, et al. "Model unlearning via sparse autoencoder subspace guided projections." Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

[3] Shen, William, et al. "LLM unlearning via neural activation redirection." Advances in Neural Information Processing Systems 38 (2026): 44253-44290.

Questions for authors

What is the overhead of constructing the aggregate forgetting subspace?

Why can the Disentangle-Head provide a relative forget–retain optimization signal? How does it change the model? Could the authors show what the head actually does to the forget representations, for example, through the ranking loss curve or an experiment in which the head is frozen?

Why are other representation-based methods discussed but not compared experimentally? What are the benefits of ERASER over LUNAR [3]?

In Table 2, ERASER achieves 0.566 on MMLU and 0.248 on WMDP-Cyber, whereas Table 12 reports 0.5679 and 0.2436, respectively, for ERASER with the head. Why are different results reported for the same method on the same model? This discrepancy raises concerns about the reported experimental results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.