acceptodds
Under review as a conference paper at ICLR 2027

Forget Without Seeing: Unlearning Behaviour Transfer without Access to the Forget Set

Abstract

Transferring unlearning ability to personalized models poses a practical dilemma: delivering the data to be forgotten exposes them anew, while directly replacing the downstream model with the upstream unlearned model or applying the upstream parameter update may disrupt locally acquired capabilities. We study a Forget Without Seeing problem, in which downstream models inherit upstream unlearning behaviour without receiving the original forget set during the unlearning update. We propose Carrier-mediated Unlearning Behaviour Transfer (CUT), inspired by subliminal learning. CUT encodes the behavioural change induced by upstream unlearning into a carrier of synthetic non-target text, enabling personalized model owners to learn endpoint-specific updates without accessing the forget set. CUT consists of three stages. First, a Teacher-Matched Steering Vector is learned to align the Original Model with the Unlearned Teacher on forget and retain calibration data. Second, the same Original Model generates paired responses to synthetic non-target prompts with and without steering, forming a Paired-Text Carrier. Finally, each downstream model learns an endpoint-specific update from the carrier, transferring unlearning behaviour while preserving personalization. Only the text carrier is delivered downstream; the forget/retain data, Unlearned Teacher, Teacher logits, and steering vector remain upstream. We evaluate unlearning transfer on TOFU across Qwen2.5-7B, Gemma-7B, and Llama-2-7B, and personalization preservation on LaMP-3. CUT transfers upstream unlearning behaviour across algorithms and forget-set sizes. On Qwen2.5-7B with SimNPO, it achieves unlearning comparable to directly applying the upstream parameter difference while reducing personalization mean absolute error (MAE) by approximately 20%. Privacy evaluations further show that the text carrier limits target-information exposure under the tested auditing, reconstruction, and update-role inference attacks. These results demonstrate that unlearning behaviour can be transferred without redistributing the original forget data, while preserving downstream personalization and limiting information exposure during delivery.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Machine Unlearning, Large Language Models, Privacy, Activation Steering

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_aYdGOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Carrier-mediated Unlearning Behaviour Transfer (CUT) and "Forget without Seeing", bringing concepts from subliminal learning to the practice of machine unlearning and performing unlearning on personalised language models from an unlearned teacher. The CUT method uses the difference between a model and its unlearned variant to learn a steering control vector and to generate a synthetic training dataset for downstream student models which induces the desired unlearning while maintaining any personalisation adjustments it has undergone. CUT is evaluated against 3 research questions, which discuss how well unlearning is transferred, how much personalisation is retained, and compares CUT against a strong TA baseline. Furthermore, the method has been analysed for data leakage.

Overall this paper makes a novel contribution by bringing subliminal learning to machine unlearning, which is an interesting and promising direction for future research. However, I have substantial questions/clarifications before I can strongly recommend acceptance - and given sufficient discussion in a response from the authors, I would be very open to increasing my score.

Critical strengths
  • The problem is well motivated, and well presented with good structuring. Forget without Seeing and CUT are well-positioned to be useful in practical settings.
  • The CUT procedure is intuitively introduced and well defined, with a clear structure and rationale as to why it works. Using a directional prior before optimising the vector is a smart and interesting way to bias the learning.
  • Experiments and ablations are thorough and cover many bases and are well presented, clearly describing and outlining what each experiment contributes to the justification of CUT.
  • The combination of subliminal learning and steering vector learning applied to machine unlearning is an interesting direction to study, with large potential for both fields.
  • Performing experiments where the unlearned teacher was produced through a variety of methods was very appreciated, and strengthened the results greatly. Using a retrained model as the "unlearned" teacher would be a good additional experiment but is not necessary.
Critical weaknesses
  • The results section is lacking the forget quality metric of TOFU, which is one of the main metrics used to analyse how well a model has forgotten something. Additionally, experiments are reported over 3 seeds, but no statistical significance is mentioned or described. Finally, the raw-p metric is not defined in the main text making it difficult to interpret the results.
  • The comparison to Task Arithmetic is interesting because TA is a much simpler method, which also doesn't require the forget set. The paper would benefit from a deeper dig into this - how well does TA retain personalisation relative to CUT?
  • Discussion of unlearning transfer and personalization preservation deserve separate analysis and discussion as this was difficult to untangle. The evaluation tasks (TOFU, and LaMP-3) would also benefit from brief discussion and why they were chosen.
  • In section 5.3.2 it is not clear how much information about the forget set is contained with the steering vector and if the steering vector can be recovered from the constructed pairs
  • The provided code does not contain trained model checkpoints, dataset split, the optimisation schedule for the steering vector, or complete training recipes nor metric code. This means I cannot run experiments to independently verify your results. Any attempt by me to reproduce your results would not be possible here - could the authors please provide this?
  • It is not immediately clear in the main text how the prompts are synthetically constructed, and the appendix discussion does not clarify why they are constructed like this. The main text would benefit from some discussion on this.
  • A concern that comes to mind with such a technique for unlearning is it's reversibility - if someone took all the PTCs and flipped chosen and rejected, then trained on this, would this recover the unlearned behaviour?
  • A very mild point is that this work is limited to language model evaluations, and this limitation is not discussed by the authors and the title mis-advertises the work as general unlearning.
Questions for authors
  • In this study, the student models are of the same architecture as the teacher model - does CUT generalise to distillation between non-uniform architecture student-teacher setups?
  • How does the setting of "Forget without Seeing" differ from others which study unlearning without access to the forget set such as PBU?
  • It is not clear how large the generated dataset corpus must be in order to use CUT, how many paired-text carriers must be produced?
17 comments
AuthorsOpenReview ↗

Author response to Reviewer aYdG (1/6): W1–W2

We thank the reviewer for the constructive feedback. We address the concerns and questions point by point below. References to our manuscript use the original submission’s numbering unless the revised manuscript is explicitly specified.

W1: Forget Quality, statistical significance, and the definition of raw-P

We address Forget Quality (FQ), statistical significance, and the definition of raw-P separately below.

1. Additional FQ results and analysis. We added FQ evaluations on Qwen2.5-7B with TOFU forget01 using the same eight Personalized Endpoints as the main experiments. Recompiled Student results follow the main-experiment endpoint aggregation and are averaged across three training seeds; see Appendix I.1, Table A25 of the revised manuscript:

Upstream unlearning algorithmUnlearned Teacher FQ ↑Recompiled Student FQ ↑
SimNPO0.02860.5830
NPO0.26570.8169
GD0.05410.0298

We also observed and analysed an interesting pattern: some Recompiled Students achieve higher FQ than their corresponding Unlearned Teachers. FQ evaluates whether the distribution of relative probabilities assigned to fact-preserving paraphrases of correct answers versus perturbed incorrect answers matches that of the retain-only reference. It does not simply measure whether the probability of the original answer decreases. Even after substantially reducing that probability, a model can still favour paraphrased correct answers over incorrect candidates more strongly than the retain-only reference, resulting in low FQ. We observe this pattern in some Unlearned Teachers, while their corresponding Recompiled Students achieve higher FQ.

Drawing on subliminal learning, we offer a possible explanation. CUT repeatedly transmits the same implicit direction through diverse fictional QA, providing a consistent training signal across examples. This may extend the update from local suppression of specific answers to changes in features shared across different formulations. Such changes could reduce the excessive preference for paraphrased correct answers over incorrect candidates, bringing the distribution closer to the retain-only reference.

With both SimNPO and NPO, the Recompiled Student achieves higher FQ than the Unlearned Teacher. In these two settings, CUT therefore not only transfers target-answer suppression but also improves unlearning quality as measured by TOFU. We will further investigate the mechanism behind this observation and the conditions under which it holds.

2. Statistical significance. Standard deviations have been added to Tables A24 and A6 in the revised manuscript’s appendix.

3. Definition of raw-P. We apologize that this point was not sufficiently highlighted in the original manuscript. We would like to clarify that Section 5.1 (page 6) already describes raw-P as reference-answer probability, and Appendix C.2.1 (page 21, Equation A20) provides the complete definition. Specifically, we compute the geometric mean of the conditional token probabilities for each reference answer, then take the arithmetic mean across QA examples. This measures the model’s propensity to predict the reference answers while reducing the influence of answer length on comparisons. In the revised manuscript, the complete definition is in Appendix D.2.1, Equation A20.


W2: Personalization preservation compared with Task Arithmetic

We apologize that this point was not sufficiently highlighted in the original manuscript. We would like to clarify that the comparison of CUT and TA in terms of personalization preservation was already reported in Section 5.2.2 (page 7), Figure 3, and Appendix Table A4. We restate the LaMP-3 results below to make this comparison explicit. Both methods use the same upstream SimNPO Unlearned Teacher and the same eight Personalized Endpoints, at the original deployment settings:

MethodMAE ↓RMSE ↓Accuracy ↑
TA0.339290.652020.72619
CUT0.267860.579180.78571

CUT reduces MAE and RMSE by 21.05% and 11.17%, respectively, relative to TA, and improves accuracy by 5.95 percentage points. CUT’s MAE is also close to the pre-update value of 0.26190, demonstrating better preservation of existing personalization while transferring unlearning behaviour. The corresponding results are in Table A6 of the revised manuscript.


AuthorsOpenReview ↗

Author response to Reviewer aYdG (2/6): W3–W4

W3: Separating unlearning transfer from personalization preservation and explaining the benchmarks

We apologize that the original manuscript briefly introduces the benchmarks and their evaluation roles in Section 5.1 (page 6) but does not explain them sufficiently clearly. We provide a more detailed description of the benchmarks and the rationale for their selection below. TOFU is a synthetic QA benchmark for LLM unlearning, containing 200 fictional authors with 20 QA pairs per author, for a total of 4,000 pairs. The questions cover biographical knowledge such as birth background, family relationships, writing genres, works, and awards. The forget set contains QA about designated authors, while the retain set contains QA about the remaining authors. TOFU also provides Real Authors and World Facts to evaluate retention of real-author knowledge and general factual knowledge. Fictional entities make the source of target knowledge and the scope of forgetting explicit and controllable, making TOFU suitable for evaluating entity-specific unlearning behaviour. TOFU

LaMP-3 is the personalized product-rating prediction task in the LaMP benchmark, built from Amazon product reviews. Each example contains a review whose rating must be predicted, its 1–5-star rating, and a user profile consisting of historical reviews and ratings. The task requires the model to use this history to capture the user’s rating habits and preferences and predict the rating of the current review. The benchmark provides both user-based and time-based splits to evaluate personalization for new users and prediction of future interactions for existing users, respectively. MAE and RMSE measure rating prediction error. LaMP

TOFU and LaMP are established, widely used benchmarks for LLM unlearning and personalization, respectively, and have been adopted in subsequent work such as OpenUnlearning and OPPU. Their clear task definitions and established metrics allow us to evaluate the two objectives separately: TOFU tests whether upstream unlearning behaviour transfers, while LaMP-3 tests whether existing personalization is preserved.


W4: Information encoded in the steering vector and its recoverability

We apologize that this point was not sufficiently highlighted in the original manuscript. We would like to clarify that steering-vector recovery and the corresponding privacy attacks were already evaluated in Sections 5.3.2 and 5.3.3 (page 8), Appendix E.2.1 (page 27), and Appendix E.4.1 (page 31). We summarize the methods and results below to make both recoverability and forget-set information exposure explicit.

The steering vector can be approximately recovered from the Paired-Text Carrier. We freeze the model parameters and use only the carrier’s prompt/chosen text to learn an intervention direction, amplitude, and layer window by minimizing cross-entropy on the chosen completions. The recovered direction has a cosine similarity of 0.84785 with the true direction; see Appendix E.2.1.

We use this recovered vector in the privacy attacks reported in the main text. Content inversion in Section 5.3.2 recovers no complete target name, predefined factual keyword, or complete answer, with a maximum answer ROUGE-L F1 of 0.231. Update-role inference in Section 5.3.3 achieves an AUROC of 0.563. These results show that the recovered vector contains some forget-set information, but the tested attacks do not reconstruct original forget examples and only weakly distinguish forget from retain roles.

We further evaluate a generative attack with stronger access by providing the true vector, which remains upstream and is not delivered downstream under the normal protocol. Unlike the main-text inversion attack, which searches for content through iterative token replacement, this attack guides generation using token-probability differences between the pre- and post-update models and those induced by vector interventions. Each condition generates 160 outputs. Maximum answer ROUGE-L F1 is 0.25926 with the recovered vector and 0.38889 with the true vector. Neither condition recovers a complete target name or complete answer; see Appendix E.4.1 and Table A13.

Access to the true vector increases textual overlap but still does not recover complete target content. These direct attacks show that the steering vector retains unlearning-related behavioural signals and is recoverable, but those signals do not translate into disclosure of complete target identities or original answers under the tested attacks. The corresponding revised locations are Appendices F.2.1 and F.4.1, including Table A22.


AuthorsOpenReview ↗

Author response to Reviewer aYdG (3/6): W5–W6

W5: Code and reproducibility materials

We have supplemented the anonymous code repository with model checkpoints, fixed data splits, the steering-vector optimization schedule, training configurations, and metric computation code.


W6: Synthetic prompt construction and its rationale

We apologize that this point was not sufficiently highlighted in the original manuscript. We would like to clarify that Section 4.2 (page 5) already describes the carrier-generation pipeline, while Appendix B.2 and Table A1 (pages 15–16) provide the complete prompt templates, sampling rules, and examples. We explain the design rationale and construction process more explicitly below.

Design basis. Our prompt construction draws on the template-based, randomized construction of synthetic data in Subliminal Learning and Subliminal Steering. The former uses random numerical prefixes and code templates to construct generation tasks. The latter programmatically combines context, generation instructions, quantity requirements, and output formats from fixed template pools. We adapt this approach to fictional-author QA unrelated to the forget targets. Subliminal Learning, Subliminal Steering, Appendix A.2.

How prompts are constructed. We programmatically combine randomly generated fictional-author names with predefined question templates. First and last names are sampled separately from fixed name lists. Questions are drawn from ten categories of biographical QA templates covering writing genres, parents’ occupations, works, awards, birthplaces, writing styles, and related topics. Each prompt requests an answer of approximately 55–85 words that naturally incorporates two randomly sampled neutral words. It explicitly requires the experiences, book titles, places, and awards to be fictional. For example:

Elian Iver is an entirely fictional author. In approximately 72 words, answer: “What language does he usually write in, and why?” Invent the relevant experiences and naturally include pressure and durability. Do not imitate real people or existing works.

Why we use this construction. New names and fictional settings keep the prompts separate from the original forget entities and QA. Retaining a biographical QA format similar to TOFU lets us test a focused question: can non-target text alone transfer unlearning behaviour downstream without providing target examples? The ten question categories cover different QA content, while the requested word count controls the expected answer length. Neutral words introduce content variation within each template. Beyond changing author names, we require answers to incorporate randomly sampled words naturally, encouraging different narrative material and phrasing within the same question type. This is intended to reduce dependence on fixed response patterns. The chosen and rejected completions in each pair use identical lexical constraints, so these constraints introduce no additional difference between the two generation conditions.

How prompts produce a training signal. For each prompt, the same Original Model generates chosen and rejected completions with and without steering, respectively, using matched randomness to reduce sampling noise. These labels indicate generation conditions, not the factual correctness of the fictional answers. The downstream model learns the steering-induced behavioural change from these pairs without receiving the target authors’ original QA. Validity filtering and signal selection that preserves category coverage produce a Paired-Text Carrier containing 3,072 pairs. Only the three text fields—prompt, chosen, and rejected—are delivered. We have expanded Section 4.2 to explain the prompt form and design rationale; the detailed construction is in Appendix C.2 and Table A2 of the revised manuscript.


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

The paper studies unlearning without access to the original forget set. Inspired by subliminal learning, the authors argue that unlearning behaviour can be transferred from a pre-trained model to its fine-tuned models. The authors propose CUT to realize unlearning transfer on TOFU from pre-trained models to the corresponding personalized models.

Critical strengths
  1. The paper has a clear motivation. The connection between subliminal learning and unlearning transfer is well stated.
  2. The experiments cover three model families.
Critical weaknesses
  1. The proposed method is complex and requires tuning for many hyperparameters (10 loss coefficients across stages).
  2. The evaluation is narrow. Unlearning is evaluated only on TOFU, and the full set of metrics is reported only for Qwen2.5-7B on forget01. On forget02/forget03 and Gemma-7B/Llama2-7B experiments, there is no evaluation on Real Authors/World Facts or personalization of the unlearned student models. There are no experiments on NPO and GD for Gemma-7B/Llama2-7B, and Qwen2.5-7B on forget02/forget03.
  3. The optimization of the Teacher-Matched Steering Vector is complex with 7 loss coefficients, and not well motivated. The authors argue that using the steered model for paired-text curation is more effective than directly using the unlearned teacher model (supporting evidence in Table 5). However, this comparison is unfair because the experiments are conducted using the set of hyperparameters that is tuned for the steered model.
  4. The proposed method relies on synthetic prompt curation with heuristic rules specified only for TOFU. It is unclear how to extend this idea to other unlearning benchmarks (e.g., WMDP, MUSE). Moreover, the template for prompt curation mirrors TOFU's own question. Therefore, the carrier is in the same domain as TOFU, which significantly weakens the subliminal framing.
  5. No computational cost is reported. For new unlearning requests, CUT requires a new steering vector optimization and paired-text generation before unlearning, which makes it computationally expensive. In contrast, TA needs weight addition only.
  6. The paper's presentation is not clear. Some details are missing, e.g., the tuning process and values of loss coefficients used. There are also inconsistencies, e.g., the abstract states 20% reduction while Section 5.2.2 states 21%. There are many other issues; the authors should revise the manuscript themselves.
Questions for authors
  1. Can you systematically report the full set of metrics for forget01/forget02/forget03 experiments on Gemma-7B/Llama2-7B/Qwen2.5-7B with SimNPO/NPO/GD methods?
  2. How does CUT perform on other unlearning benchmarks such as WMDP and MUSE? How to construct the synthetic prompts for these domains?
  3. Can CUT achieve unlearning using synthetic prompts whose domain is different from that of the unlearning domains?
  4. Can you provide details of hyperparameter tuning for CUT? Can you provide values for all hyperparameters used in the paper?
  5. How does replacing the steering vector vv with simpler ones affect the overall performance? The simpler ones could be the vector dd in Eqn. 3 or the difference-in-means between forget representations of the unlearned teacher and original models.
  6. How does Task Arithmetic perform across a range of coefficients?
  7. Can you report the performance of applying unlearning methods on personalized models directly? This can serve as the oracle baselines for comparison.
15 comments
AuthorsOpenReview ↗

Author response to Reviewer HSoc (1/7): W1, Q4, W3, Q5

We thank the reviewer for the feedback. Below, we address the reviewer’s concerns and questions point by point.

W1 & Q4 & W3 & Q5: Steering design, signal strength, and hyperparameters

Regarding the complexity concerns in W1 and W3, we agree that the current design uses many loss coefficients and could be simpler and more elegant. This complexity reflects specific requirements of upstream signal encoding: intervention direction, teacher behaviour matching, signal strength, and non-target knowledge retention must be considered jointly. Section 4.1 organizes vector optimization into directional regularization, forget/retain teacher matching, and probability calibration. These objectives address the corresponding requirements so that the strengthened signal supports selective unlearning.

This additional complexity yields clear practical gains. Following Q5, we added simple vector-replacement experiments on Qwen2.5-7B with TOFU forget01 and SimNPO, matching vector norms, carrier prompts, and the Student training configuration. Aggregated results across three Personalized Endpoints are shown below:

Vector constructionForget raw-P ↓Retain raw-P ↑
Forget/retain representation-difference direction0.317600.25946
Forget-only mean representation-difference direction0.421320.31546
CUT: Teacher-Matched Steering0.071270.58157

Teacher-matched optimization improves both target suppression and non-target knowledge retention, demonstrating the practical value of the additional optimization.

Regarding the fairness concern in the second part of W3, we would like to clarify that our primary criterion for choosing the carrier-generation route is signal strength, which is key to effective transfer. Section 4.1 explains this rationale, and Appendix C.2 reports the paired preference signals provided by both routes before Student training. Results for their respective 1,024-pair high-signal subsets are shown below:

Carrier-generation routeMean preference contrast ↑
Direct generation with the Unlearned Teacher1.76203
Steered Original3.54801

The Steered Original provides approximately 2.01 times the mean preference contrast of direct Teacher generation. This upstream evidence is the main basis for choosing the steering-based generation route, independently of subsequent Student hyperparameter selection. The signal difference exists before Student training and is therefore not a result of subsequent Student hyperparameter tuning.

Regarding Q4, we have consolidated the tuning procedure and hyperparameter values in Appendices C.1 and C.3, including the loss coefficients, learning rates, and update counts for the four stages of vector optimization, as well as the key Student training settings.

AuthorsOpenReview ↗

Author response to Reviewer HSoc (2/7): W2, Q1

W2 & Q1: Evaluation coverage and complete metrics

We apologize for the incomplete metric reporting in some extension experiments. We now provide the missing Forget, Retain, Real Authors, and World Facts metrics.

We understand the request for broader evaluation. Given the time and computational cost, we add NPO and GD experiments on Gemma-7B and Llama-2-7B with forget01, extending coverage across model families and unlearning algorithms. For changes in forget-set size, we retain the existing forget02/forget03 observations with SimNPO, which achieves the best overall performance in our main experiments. Full results are reported in Appendix E.2.

AuthorsOpenReview ↗

Author response to Reviewer HSoc (3/7): W4, Q2, Q3

W4 & Q2 & Q3: Multi-domain carriers and transfer beyond TOFU

Thank you for this suggestion. Fictional-author QA is one carrier format we use; CUT does not require it. To examine dependence on prompt format and domain, we added multi-domain carrier experiments and extended the evaluation to MUSE-News.

For W4/Q3, we construct synthetic passages spanning 12 domains, including fictional authors, engineering, software, ecology, materials, and transportation. We use a shared rewriting instruction that preserves facts and entities. For each prompt, we generate paired responses from the same Original Model with steering enabled and disabled. Filtering for text quality and overlap with forget content produces a Paired-Text Carrier containing 3,072 response pairs. Steering encodes the unlearning signal in the paired responses; the prompts need not follow TOFU’s author-QA templates or explicitly specify the forget targets.

Results on Qwen2.5-7B with TOFU forget01 and SimNPO are shown below:

Model/carrierForget 1−ROUGE-L ↑Forget 1−raw-P ↑Retain ROUGE-L ↑Retain raw-P ↑Real Authors ROUGE-L ↑Real Authors Norm. P ↑World Facts ROUGE-L ↑World Facts Norm. P ↑
Personalized Endpoint0.126140.047960.871370.944560.800000.455020.900280.46483
CUT, 12-domain carrier0.719490.923280.456770.543500.626000.434700.803700.46004

Mean personalization metrics across three Personalized Endpoints are shown below:

Model/carrierMAE ↓RMSE ↓Accuracy ↑
Personalized Endpoint0.269840.562420.77778
CUT, 12-domain carrier0.285720.604800.76190

These results show that CUT transfers unlearning behaviour through multi-domain rewritten text while largely preserving local personalization.

For W4/Q2, we further evaluate CUT on Llama-2-7B with MUSE-News and SimNPO. We reuse the same 12-domain prompt design, optimize the Teacher-Matched Steering Vector using news forget/retain data, and generate paired text. No news-specific QA templates are required.

Aggregated results across three Personalized Endpoints are shown below, with KnowMem metrics expressed as percentages:

ModelKnowMem-F ↓KnowMem-R ↑
Personalized Endpoint56.2249.44
Unlearned Teacher36.8641.37
CUT, 12-domain carrier39.4543.22

The corresponding mean personalization metrics are:

ModelMAE ↓RMSE ↓Accuracy ↑
Personalized Endpoint0.238100.483410.77778
CUT, 12-domain carrier0.253970.520980.76190

CUT achieves KnowMem-F close to the Teacher’s and higher KnowMem-R, while personalization accuracy decreases by approximately 1.59 percentage points. These results demonstrate that multi-domain carriers transfer unlearning behaviour on the news benchmark.

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

The paper studies transferring upstream unlearning behaviour to already-personalized downstream models without delivering the original forget set. CUT learns a residual steering vector to match an unlearned teacher on forget and retain calibration data, generates paired synthetic responses with and without steering, and trains a downstream LoRA through endpoint-relative DPO with suppression and personalization safeguards. Experiments evaluate behavioural transfer on TOFU across three model families, personalization on LaMP-3, and information exposure under selected attacks. The upstream party retains access to the forget/retain data; the downstream update receives only the text carrier.

Critical strengths
  1. The deployment problem is useful and clearly motivated. Delivering an unlearning update while retaining local personalization and avoiding redistribution of forget data is a meaningful objective. The paper specifies the upstream/downstream information boundary and acknowledges that recipient-side data blindness does not make the entire pipeline data-free.

  2. The experiments provide a credible feasibility signal within the evaluated same-base setting. Target-answer suppression is substantial, and the random-direction, direct-teacher-carrier, and loss ablations show that the proposed components matter. The direction-alignment analysis and personalization results also provide useful evidence about what is transferred and what is preserved.

Critical weaknesses
  1. The demonstrated increment over closely related work is limited. Behavioural transfer through text that omits the target trait is already established by Cloud et al., "Subliminal Learning." Morgulis and Hewitt, "Subliminal Steering," already train steering vectors, transmit their effects through generated text, and recover aligned directions. Blank et al., "Subliminal Learning Is Steering Vector Distillation," further study this mechanism through direction extraction and causal interventions. CUT cites these works, which I appreciate; the concern is the size of the advance. Its clearest additions are the unlearning/personalization setting, teacher-matched vector construction, paired-text DPO, and protection terms. These could form a useful method, but the current evidence does not sufficiently establish a substantial new capability beyond applying the existing mechanism to this setting. The personalization advantage over Task Arithmetic also needs comparison at calibrated, comparable forgetting/retention points, since that advantage is central to the practical contribution.

  2. Target-specific transfer and its dependence on shared model ancestry remain insufficiently identified. Table 1 shows stronger suppression on forget than retain examples: under SimNPO, forget raw-P falls from 0.94819 to 0.07987, whereas retain raw-P falls from 0.93335 to 0.58226. This is evidence of differential suppression. However, retain ROUGE-L also falls from 0.84011 to 0.49739, and the carrier remains in the same fictional-author biography domain. The random-vector control does not distinguish target-identity transfer from a learned, domain-specific confidence or response shift. Appendix D.3 reports a cosine similarity of 0.9917 between forget-derived and retain-derived student directions and correctly acknowledges that alignment alone does not establish specificity. A common direction can still have selective output effects; the missing evidence is a controlled test showing that changing the forget identities changes which identities are suppressed.

The downstream model is explicitly adapted from the same original model, and the other model families use independently constructed teachers, vectors, and carriers. These experiments demonstrate repeated within-family feasibility, but do not quantify how much personalization or representational drift the same carrier can tolerate. Shared initialization is important in prior work's local distillation theory; its guarantee does not directly cover personalized students trained on filtered, sampled text through LoRA/DPO. The paper provides neither a corresponding analysis nor a controlled compatibility study. Together, these gaps leave the scope of the proposed selective transfer across personalized endpoints unclear and currently outweigh the strengths.

Questions for authors
  1. Clarification of the contribution and comparison. What capability is added beyond the three closely related works above, and which design element is essential for it? Please distinguish the new deployment setting from a new transfer mechanism. For the practical advantage over Task Arithmetic, please allow both methods upstream calibration under the same information boundary and fixed selection rules, then compare retention and personalization at comparable forgetting strengths, using both answer-probability and generation metrics.

  2. Requested validation of target specificity. Could the authors run a small cross-target experiment using two disjoint, difficulty-matched forget sets F1 and F2, a fixed original model, and the same personalized endpoints? Construct carriers C1 and C2 with matched prompts and budgets, then evaluate both resulting students on both forget sets and a common held-out retain set that excludes their union. Report the matrix D_ij = mean_{(x,y) in Fj}[NLL(M_i; x,y) - NLL(M_P; x,y)], where M_i is trained on Ci and NLL is measured per completion token, alongside generation and retain metrics. Preferential diagonal suppression would show that the carrier conveys which targets to suppress. Please also include a generic confidence-reduction or biography-domain suppression control calibrated on separate data, and compare at matched target suppression or matched retain damage. Uncertainty should account for shared authors and endpoints. These controls would directly address the main alternative explanation within the existing benchmark.

  3. Clarification and bounded validation of compatibility. Is shared model ancestry a stipulated deployment restriction, and what departures from the original model remain supported? Please hold a carrier fixed while varying personalization strength on the existing base model, and relate transfer, retain utility, and local-task preservation to a quantitative compatibility measure, such as directional Jacobian or carrier-induced gradient alignment. A same-architecture, different-initialization control could provide an additional diagnostic if feasible. Please distinguish post-training activation alignment from a condition that predicts successful transfer. As an optional analytical extension, a local sufficient condition connecting carrier observability, teacher-student compatibility, and selective forget/retain changes would be useful; clearly state which parts extend to the actual steering/text/LoRA/DPO pipeline. The priority is a measured applicability boundary and the target-specific controls above, rather than a universal theorem for arbitrary networks.

9 comments
AuthorsOpenReview ↗

Response to Reviewer ZcYh (1/7): W1 & Q1 — Unlearning transfer designs

We thank the reviewer for the careful review and suggestions. We address the concerns and questions point by point below.

W1 & Q1: CUT’s added capability, key designs, and comparison with calibrated TA

  1. CUT’s added capability and key designs

Thank you for asking about CUT’s added capability and key designs. Our work is inspired by the three studies cited above. These studies established that hidden behaviours can transfer through text and identified the key role of steering vectors, primarily for relatively simple behavioural targets such as preferences and phrases. We develop this mechanism for privacy-preserving updates and extend it to practical unlearning needs. CUT transfers complex unlearning behaviour from upstream models to already-personalized downstream models while limiting target-information exposure and preserving local capabilities.

This extension requires jointly addressing three new questions. Our methodological additions and key designs address these questions.

(1) Can complex unlearning behaviour transfer effectively through non-target text?

Unlearning requires selectively suppressing specified targets while preserving non-target knowledge. This composite behaviour is distributed across high-dimensional parameters and may not produce a sufficiently strong, consistent learning signal in non-target text. We address this challenge with four complementary designs.

First, we introduce joint optimization of the Teacher-Matched Steering Vector for selective unlearning. We combine a forget/retain representation-difference prior, teacher matching on both sets, and calibration of behavioural strength so that the vector encodes both target suppression and non-target retention.

Following another reviewer’s suggestion, we added simple vector-replacement experiments comparing two directly extracted directions. The forget/retain representation-difference direction subtracts the mean representation change on retain examples from that on forget examples, with both changes measured from the Original Model to the Unlearned Teacher. The forget-only mean representation-difference direction uses only the mean representation difference between these models on forget examples. Neither baseline undergoes joint vector optimization. Results across three Personalized Endpoints are shown below:

Vector constructionForget raw-P ↓Retain raw-P ↑
Forget/retain representation-difference direction0.317600.25946
Forget-only mean representation-difference direction0.421320.31546
CUT: Teacher-Matched Steering0.071270.58157

Joint optimization improves both target suppression and non-target knowledge retention, demonstrating the practical value of optimizing the vector specifically for unlearning behaviour.

Second, we generate the carrier with the Steered Original rather than directly with the Unlearned Teacher. We apply the Teacher-Matched Steering Vector to the Original Model, encoding the Teacher’s behavioural change as a controllable residual intervention that strengthens the unlearning signal in synthetic non-target text. On Qwen2.5-7B with TOFU forget01 and SimNPO, we compare Steered/Control Original pairs with Teacher/Original pairs while keeping Student training unchanged. Our generation route reduces Forget raw-P from 0.43811 to 0.07987, an approximately 81.8% reduction.

Third, we introduce the Paired-Text Carrier and a corresponding endpoint-relative DPO objective. For each prompt, we use the same Original Model and matched randomness to generate responses with and without steering, forming chosen/rejected pairs. Each Student uses its own Personalized Endpoint as a frozen reference and learns the intervention-induced difference through the adaptive preference loss. A rejected-probability floor controls damage to non-target knowledge. Removing the DPO term increases Forget raw-P from 0.07987 to 0.95083, effectively eliminating unlearning transfer and confirming the critical role of the paired signal and its corresponding learning objective.

AuthorsOpenReview ↗

Response to Reviewer ZcYh (2/7): W1 & Q1 — Carrier design, privacy and personalization

Fourth, we design carrier construction and selection to balance content coverage and signal strength. Ten categories of question templates, randomly generated fictional names, and neutral-word constraints increase content diversity. We then select behavioural signals specifically introduced by steering: we compute the relative preference between chosen and rejected responses under the steered and control conditions, and use the difference to quantify the additional signal induced by the intervention. The final Paired-Text Carrier combines 2,048 base pairs and 1,024 category-balanced high-signal pairs, for 3,072 pairs in total. The base pairs preserve content coverage, the high-signal pairs strengthen learnable behavioural differences, and category balancing prevents selection from concentrating on a few templates. Together, they provide a training carrier with broad coverage and a concentrated behavioural signal.

(2) Can the carrier combine effective transfer with delivery privacy?

We design the carrier for delivery privacy while strengthening the transfer signal. Fictional content, filtering of both generation streams, and text-only delivery limit target-information exposure. We conducted text auditing, unknown-content reconstruction and inversion attacks, and update-role inference given complete candidate answers. The reconstruction and inversion attacks recovered no complete target name or answer. The carrier also yielded a lower update-role inference AUROC than Gradient Delivery and the PerTA-fisher parameter patch. These results demonstrate the carrier’s ability to protect target information while effectively transferring behaviour.

(3) Can transfer preserve existing downstream personalization?

We use the update magnitudes of the existing personalization LoRA to apply module-weighted protection to the new LoRA, without replaying local data. In the Student-objective ablation, this reduces LaMP-3 MAE from 0.34524 without protection to 0.26786, close to the pre-update value of 0.26190, effectively preserving local personalized capabilities.

Our core contribution is therefore to introduce a practically valuable unlearning-update setting, substantially advance the transfer of complex unlearning behaviour through subliminal learning, and develop systematic methods and attack evaluations for security and privacy protection. We believe these additional designs and systematic experiments together constitute substantial methodological and empirical contributions. We also appreciate the reviewer’s interest in further capabilities, which we will continue to explore.

AuthorsOpenReview ↗

Response to Reviewer ZcYh (3/7): W1 & Q1 — Calibrated Task Arithmetic comparison

  1. TA coefficient sweep and downstream unlearning and personalization

On Qwen2.5-7B with TOFU forget01 and SimNPO, we report an upstream sweep over eight nonzero TA coefficients. We further report downstream unlearning, non-target knowledge retention, and personalization at three representative operating points: 0.75, 1.00, and 1.30. CUT keeps its fixed configuration from the original manuscript.

The upstream sweep results are as follows:

TA coefficientF 1−RL ↑F 1−Raw P ↑R RL ↑R Raw P ↑RA RL ↑RA Norm. P ↑WF RL ↑WF Norm. P ↑
0.500.454250.479720.795190.923670.758330.427090.888890.44535
0.750.555720.902840.589800.672750.693330.425140.857480.43077
1.000.679510.954580.542400.600180.660170.418450.839030.42820
1.300.690060.963300.495310.564410.620330.383050.815100.38952
1.600.719810.972600.399780.295780.531330.346910.766670.36269
2.000.746980.971920.366750.183030.491670.329440.773500.35292
2.500.821890.970800.276680.096560.370670.315000.740030.33816
3.000.880810.977490.178370.056840.160000.314630.702990.36066

Increasing the coefficient generally strengthens forgetting while increasing damage to non-target knowledge. Based on the upstream forgetting/retention trade-off, the coefficient of 1.00 used in the original manuscript provides a favourable overall balance. To avoid relying on a single definition of the “optimal” operating point, we also include 0.75 and 1.30 and compare downstream unlearning and personalization preservation across coefficients.

Downstream unlearning and knowledge retention for the three TA operating points and CUT are shown below:

MethodF 1−RL ↑F 1−Raw P ↑R RL ↑R Raw P ↑RA RL ↑RA Norm. P ↑WF RL ↑WF Norm. P ↑
TA, 0.750.567360.870160.587340.672540.673460.417920.842130.43200
TA, 1.000.641840.939510.522030.615700.658630.408030.833440.41851
TA, 1.300.668480.956470.510960.534280.633080.392210.821810.39928
CUT0.700750.920130.497390.582260.689830.438560.832190.46670

Mean LaMP-3 results across eight Personalized Endpoints are as follows:

MethodMAE ↓RMSE ↓Accuracy ↑
TA, 0.750.315480.6209375.00%
TA, 1.000.339290.6520272.62%
TA, 1.300.333330.6617073.81%
CUT0.267860.5791878.57%

CUT achieves overall unlearning performance comparable to TA at the original operating point while better preserving existing personalization. CUT achieves lower MAE and RMSE and higher accuracy in the comparisons with all three TA operating points. Relative to TA at 1.00, CUT reduces MAE and RMSE by 21.05% and 11.17%, respectively, and improves accuracy by 5.95 percentage points. The full results are reported in Appendix E.4 of the revised manuscript.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response and Summary of Revisions

We thank all reviewers for their evaluation and helpful feedback. We have added experiments and analyses and clarified the manuscript. The main revisions are highlighted in blue and summarized below.

  1. Clarifying the design rationale and implementation details. We further explain the design considerations behind Teacher-Matched Steering, carrier generation, and the downstream training objectives, and provide the tuning procedure and key hyperparameters. See Sections 4.1–4.2 and Appendix C.

  2. Expanding evaluation coverage and strengthening the TA comparison. We complete metric reporting for the existing extension experiments and add NPO and GD evaluations on Gemma-7B and Llama-2-7B under forget01. We also add a TA coefficient sweep and comparisons at three representative operating points. CUT better preserves personalization at all three points. See Appendices E.2 and E.4.

  3. Improving metric definitions and result reporting. We add Forget Quality evaluation and clarify the definitions and reporting conventions for raw-P and normalized probability. See Section 5.1 and Appendices D.2 and I.1.

  4. Clarifying carrier compatibility and applicability. We add fixed-carrier experiments across personalization strengths, knowledge training stages, and training branches, together with compatibility diagnostics before the update. We also examine carrier reuse across models and carrier size under a fixed update budget. See Appendices E.5, I.2, and I.3.

  5. Reporting computational costs. We report the time and GPU memory usage of each CUT stage, TA deployment time, and delivery artifact sizes. See Appendix A.

The remaining concerns and questions are addressed individually in our responses to each reviewer. We again thank all reviewers for their time and helpful feedback.

We thank the reviewers and AC for their review and discussion. Their suggestions prompted extensive additional experiments and strengthened CUT’s methodological basis, empirical validation, and reproducibility materials.

CUT’s core contribution is to transfer upstream unlearning behaviour to downstream models after release and personalization, without redistributing the original forget data, while preserving personalization and protecting delivery privacy. We encode upstream behavioural changes in a Paired-Text Carrier of synthetic non-target text. Downstream owners use the delivered carrier to fine-tune their existing Personalized Endpoints, without receiving Unlearned Teacher weights or upstream parameter patches. This extends subliminal learning to unlearning updates that jointly require selective target suppression, non-target knowledge retention, personalization preservation, and delivery privacy. Experiments across model families, upstream unlearning algorithms, and forget-set sizes validate this mechanism’s effectiveness and applicability. We further assess delivery privacy through text auditing, content reconstruction, and update-role inference. The tested reconstruction attacks recover no complete target identity or original answer.

During discussion, we made six substantive additions:

  1. Design justification and fair comparisons. We added two simple vector-construction baselines, independently tuned the Direct-Unlearned-Teacher Carrier route with a larger Student training budget, and supplemented Task Arithmetic (TA) coefficient calibration and downstream comparisons. These controls further support the Teacher-Matched Steering Vector design.

  2. Direct tests of target specificity. We added cross-target experiments and a generic-suppression control matched for retain damage. The suppressed targets change with the upstream forget targets, demonstrating that the carrier conveys which targets to suppress.

  3. Broader evaluation and carrier controls. Our evaluation covers three model families, three unlearning algorithms, and different forget-set sizes, with additional MUSE-News and MUSE-Books experiments. A pure 11-domain carrier supports effective transfer without author/biography text, and experiments at five carrier sizes further examine data requirements.

  4. Transfer compatibility, applicability boundaries, and exploratory pre-update diagnostics. Fixed-carrier experiments examine how downstream model states affect transfer. We introduce two pre-update diagnostics, the carrier gradient cosine and local directional cosine, to assess compatibility for unlearning behaviour transfer and personalization preservation, respectively. This exploratory work provides a preliminary basis for assessing carrier suitability and potential personalization interference before full training, informing whether to use the current carrier.

  5. Deployment and reproducibility materials. We added measured computational costs and communication overhead, hyperparameter configurations across model families, and pipeline execution instructions, clarifying costs, configurations, and usage conditions.

  6. Corrections to reporting and overstatements. We take errors in parts of the manuscript, insufficiently documented endpoint selection, and occasional overstatements seriously. We have corrected the relevant content, clarified selection criteria, and revised claims exceeding the empirical evidence.

These efforts strengthen CUT’s evidence through methodological analysis, controlled comparisons, applicability evaluation, and transparent reporting. We also explicitly report behavioural reversibility, limitations of carrier reuse across architectures, and incomplete transfer on some MUSE metrics. Our experiments establish CUT’s core transfer capability. Future work will explore extending CUT beyond behavioural suppression to remove target knowledge from downstream models while keeping the forget set upstream, and examine unlearning durability across successive updates. We will also validate pre-update diagnostics, refine applicability boundaries, evaluate stronger privacy attacks and corresponding defenses, and investigate this mechanism on broader datasets and vision-language models.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.