acceptodds
Under review as a conference paper at ICLR 2027

PLACEBU: Placebo-Controlled Retain-Free LLM Unlearning

Abstract

LLM unlearning aims to remove the influence of designated training data while preserving the model's remaining capabilities, with applications to privacy, copyright compliance, the removal of sensitive knowledge, and many more. To preserve model utility after unlearning, many existing methods rely on accessible retain data, typically using a small retain set as a proxy of the much larger original training corpus. However, such data can be unavailable in practice, and a randomly sampled retain set may not provide the most informative signal for determining what should be preserved. In this paper, we show that sampled retain data are not the only source of preservation supervision: the forget set itself contains targeted information about what should remain unchanged. Each forget example pinpoints the knowledge whose training influence is to be removed, while structured perturbations around that example probe the local predictive structure, separating forget-specific effects from shared effects that should be preserved. Based on this insight, we propose PLACEBU, which uses matched target-placebo controls to estimate this boundary, to achieve unlearning without access to retain examples. Experiments on TOFU and MUSE demonstrate competitive forgetting–utility trade-offs against both retain-based and retain-free baselines. On MUSE-Books, PLACEBU reduces knowledge memorization by 84.7% relative to a strong baseline while maintaining competitive utility preservation. Our code is available at https://anonymous.4open.science/r/PLACE-96A3/.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
model bias/unfairness mitigation, safety and alignment, security and privacy

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_RuviOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposed PLACEBU to solve the utility preservation in a remain data free setting in LLM unlearning. This is a polished an empirically strong paper with an intuitive matched-control idea. However, the central theoretical assumptions are not validated, the implementation is only indirectly connected to the theory, and the "retain-free" claim is weakened by retain-side hyperarameter selection.

Critical strengths
  1. The paper introduces an interesting request-specific matched-control construction.

  2. Its empirical analysis is further strengthened by random-seed experiments and adversarial evaluations, and the overall presentation is clear and well organized.

Critical weaknesses
  1. The central claim that the forget set reveals what should remain unchanged is not well substantiated. PLACEBU defines the target contrast as a mixture of forget-specific and preservation-relevant effects and assumes that a synthetically constructed non-target relation recovers the latter. However, these components are never independently observed, and the paper neither measures the proposed control mismatch nor demonstrates that semantically different target and placebo relations induce comparable nuisance variation.

  2. The theoretical results are largely assumption-driven. Proposition 3 follows directly after assuming ∥ni−nipl∥≤ϵctrl,i\|n_i-n_i^{pl}\|\leq\epsilon_{\text{ctrl},i}. More importantly, Lemma 1 assumes that the assistant residuals approximate the abstract target and placebo contrasts, which is precisely the nontrivial connection that needs justification. The current theory does not establish that the training objectives recover these contrasts.

  3. The end-to-end retain-free status of PLACEBU is unclear. Appendix C states that “retain-side scores may be used” to select the inference coefficients and mask parameters, while Appendix E.6 does not specify whether the reported configurations were selected using such scores. If retain-side validation performance informed this selection, the complete pipeline is not strictly retain-free, even though no retain examples are used to optimize the assistants.

  4. On MUSE, PLACEBU produces substantially lower memorization than the retrained reference while also losing utility. Likewise, TOFU’s aggregate metric can reward stronger-than-retrain suppression. Distance to the retrained model should be reported across all settings, rather than treating maximal forget scores as uniformly desirable.

Questions for authors
  1. Most headline improvements are small, while uncertainty across baseline runs is not reported. PLACEBU also requires the deployed model plus multiple assistant forward passes at every decoding step.
2 comments
AuthorsOpenReview ↗

Response to Weaknesses 1 - 3

W1. validation of the matched-control assumption

We agree that the latent nuisance component and the exact control mismatch ϵctrl,i\epsilon_{\mathrm{ctrl},i} cannot be observed directly. We therefore test their observable implication: whether placebo compensation reduces deletion-induced changes on matched non-target information.

To separate compensation from overall forgetting strength, we compare the closest available MEM. operating points while fixing β−=2.1\beta_-=2.1. Deletion-only uses β+=0\beta_+=0, whereas the diagnostic compensated setting uses β+=0.25\beta_+=0.25, which gives the closest MEM. among the tested compensation strengths. This setting is used only for this analysis.

MethodMEM. ↑Non-target KL ↓Residual RMS ↓Non-target answer Δlog⁡p\Delta\log p ↑
Deletion only0.8947.2054.385-5.324
Deletion + placebo compensation0.8835.6803.832-4.158

With only a 0.0110.011 MEM. difference, compensation reduces non-target KL by 21.2% and residual RMS by 12.6%. KL decreases on all 800 controls, and RMS decreases on 799/800 controls.

The existing control ablation further shows that the matched structure matters:

Control constructionAGG. ↑MEM. ↑UTIL. ↑
Random retain control0.480.520.45
Semantic-nearest retain control0.560.500.63
Matched placebo control0.650.680.63

Together, these results show that matched placebo compensation cancels a substantial portion of deletion-induced non-target change, beyond what is obtained from random or semantic-nearest controls.


W2. Connection between the theory and assistant training

Logit-difference correction is an established unlearning mechanism: ULD derives an unlearned distribution by subtracting an assistant model’s logits, while Offset Unlearning learns an unlearning offset by contrasting the logits of smaller models. PLACEBU builds on this established mechanism; its new contribution is to obtain two separately supervised residuals corresponding to the target and matched-placebo interventions.

For each branch, the trained assistant and its frozen reference share the same initialization. The deletion branch is trained on C11C_{11} and constrained on C01C_{01}, while the compensation branch is trained on C10C_{10} and constrained on C00C_{00}. Therefore, the formula below rσ=ℓaσ−ℓa0σr^\sigma=\ell_{a^\sigma}-\ell_{a^\sigma_0} isolates the training-induced logit change associated with that intervention pair. PLACEBU subtracts the target-derived residual and restores the matched-placebo residual.

We agree that Lemma 1 is a conditional error decomposition rather than a guarantee of exact recovery: ei−e_i^- and ei+e_i^+ explicitly represent the remaining approximation errors. We have clarified this point in the updated manuscript.

Empirically, the learned compensation residual reduces non-target KL by 21.2% and residual RMS by 12.6% at a closely matched forgetting level, supporting the intended correspondence between the supervised branches and their operational roles.


W3. End-to-end retain-free status

Thank you for noting that “retain-side scores may be used” in Appendix C makes the protocol ambiguous. It refers to an optional operating-point procedure, not the protocol used for the reported PLACEBU results.

PLACEBU does not require a retain set. Retain examples are used only after configuration selection to report UTIL. They are not used for quartet construction, assistant training, checkpoint selection, or inference-parameter selection.

Calibration of β−\beta_- and β+\beta_+ is also retain-free. We first select the smallest β−\beta_- satisfying the prescribed forget-side MEM. requirement. We then initialize β+\beta_+ by matching the deletion and compensation residual scales:

β+init=β−RMS(r−)RMS(r+).\beta_{+}^{\mathrm{init}}=\beta_{-}\frac{\mathrm{RMS}(r^{-})}{\mathrm{RMS}(r^{+})}.

This normalization accounts for differences in the magnitudes of r−r^- and r+r^+. Around this initialization, we select β+\beta_+ and the support-mask parameters by minimizing residual changes on the constructed C10/C00C_{10}/C_{00} controls while maintaining the MEM. requirement. This uses only C11C_{11} forget queries and forget-derived C01C_{01}, C10C_{10}, and C00C_{00} controls. The resulting parameters are fixed before retain-side evaluation and shared across benchmark requests.

Table 22 shows an operating-point trade-off, not dependence on retain data. Increasing β−\beta_- strengthens forgetting, while increasing β+\beta_+ restores non-target behavior but may moderately weaken forgetting. All settings use the same frozen assistants, with no retain examples used in training or calibration.

We have removed the ambiguous sentence from Appendix C, added the calibration procedure, and distinguished parameter selection from post-hoc utility evaluation. Under the reported protocol, PLACEBU remains end-to-end retain-free.

AuthorsOpenReview ↗

Response to Weakness 4 & Question 1

W4. Stronger-than-retrain suppression

We agree that stronger suppression than Retrain should not automatically be interpreted as better unlearning. Our claim is that PLACEBU provides a controllable forgetting–utility trade-off without retain data, rather than that maximal forgetting is always desirable.

For example, on 8B Forget10, Retrain obtains MEM./UTIL. =0.57/0.75=0.57/0.75, while the default PLACEBU setting obtains 0.68/0.630.68/0.63, corresponding to absolute differences of 0.11/0.120.11/0.12. However, the existing coefficient sweep shows that PLACEBU can move toward the Retrain operating point: with β−=1.2\beta_-=1.2 and β+=2.0\beta_+=2.0, it obtains MEM./UTIL. =0.55/0.69=0.55/0.69, reducing the corresponding differences to 0.02/0.060.02/0.06.

In the updated manuscript, we have added the distance to Retrain alongside the direction-specific metrics for all settings and revise the text to avoid treating maximal suppression as uniformly preferable. The stronger suppression observed on MUSE, particularly for Books KnowMem, are described as a calibration limitation rather than an unconditional improvement in our revised paper.


Q1. Magnitude, uncertainty, and inference cost

We respectfully clarify that the improvements are not uniformly small. On 8B TOFU, PLACEBU achieves the highest AGG. at all three forget ratios, exceeding the strongest baseline by 0.040.04, 0.030.03, and 0.010.01 on Forget1/5/10. The Forget10 margin is modest, but the advantage is consistent across model scales and forget ratios. The gains are larger in other settings: on Qwen Forget5, PLACEBU improves AGG. from 0.550.55 to 0.630.63; on MUSE-Books, it reduces KnowMem from 30.11%30.11\% to 4.60%4.60\% relative to the utility-matched SimNPO baseline while maintaining comparable utility (60.81%60.81\% vs. 60.13%60.13\%).

PLACEBU is stable across three independent assistant-training seeds. On 8B Forget10, the standard deviations of AGG., MEM., and UTIL. are 0.0030.003, 0.0020.002, and 0.0050.005, respectively. Since some baselines are available only through reported results or single checkpoints, we cannot reliably estimate their training-seed variance without fully reproducing them. We therefore avoid interpreting small margins, such as the 0.010.01 Forget10 gain, as statistically significant and distinguish them from the larger improvements observed on MUSE.

PLACEBU follows the auxiliary-model decoding paradigm established by ULD, which also constructs an unlearned distribution through logit differences between the deployed model and a lightweight assistant. ULD explicitly acknowledges the resulting inference latency and notes that it can be mitigated by executing the deployed model and assistant in parallel.

PLACEBU extends this paradigm with separate deletion and compensation residuals. Each decoding step requires one deployed-model pass and four lightweight residual passes. However, each PLACEBU assistant contains only four Transformer layers, compared with the eight-layer assistant used by ULD. The four residual passes are independent, support KV caching, and can be batched or executed concurrently with the deployed model. Thus, four residual passes do not imply a four-fold increase in wall-clock latency, although PLACEBU does incur additional total computation and memory.

The additional computation implements the matched-control compensation that improves the forgetting–utility balance: on 8B Forget10, PLACEBU obtains AGG. =0.65=0.65, compared with 0.550.55 for ULD, without using original retain examples to train the assistants. We have revised the paper accordingly to report end-to-end decoding latency, throughput, and peak memory under both sequential and parallel execution, making this quality–efficiency trade-off explicit.

Reviewer_3m3YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PLACEBU, a retain-free LLM unlearning framework that aims to selectively remove the influence of designated training data while preserving model utility. The authors argue that existing localization-based methods cannot reliably separate forget-specific effects from preservation-relevant information, as both may coexist within the same contextualized representation. To address this limitation, PLACEBU introduces a matched 2 × 2 target–placebo design that compares an original entity and its controlled replacement under both a target relation and a non-target placebo relation. Two lightweight assistants, a deletion assistant and a compensation assistant, are trained on the corresponding intervention pairs. Their reference-relative logit residuals are then combined with the frozen deployed model's logits to achieve selective unlearning without using original retain examples to train either assistant.

Critical strengths

Theoretical Motivation for Effect Separation Beyond Localization: Proposition 1 provides a theoretical explanation of why localization alone is insufficient to distinguish forget-specific effects from preservation-relevant information. The analysis establishes that even accurate localization cannot uniquely identify the component that should be removed when both effects coexist within the same contextualized representation. This provides a meaningful motivation for moving beyond localization-based unlearning toward explicit effect separation.

Selective Unlearning Through Matched Target–Placebo Controls: Rather than relying solely on localized forget targets, PLACEBU constructs matched target–placebo controls to estimate and compensate for non-target variation introduced by entity replacement. The contrast-of-contrasts formulation offers a principled approach to separating forget-specific effects from preservation-relevant variation. Furthermore, achieving competitive forgetting–utility trade-offs without accessing original retain examples is a practically valuable contribution.

Critical weaknesses

Insufficient Direct Analysis of Collateral Effects during Unlearning: Section 2.2 explains that forget-specific information and preservation-relevant information may coexist within the same representation. Although Table 3 and Appendix F.3 provide evidence of improved utility preservation, they do not directly demonstrate how the proposed method separates these effects. A quantitative analysis of token-level logit changes would help clarify how much non-target information is affected by the deletion correction and how effectively the placebo compensation reduces these collateral effects.

Insufficient Direct Evaluation of Targeted Knowledge Suppression: The proposed method aims to selectively suppress specific entity–relation associations. Although the evaluation includes exact memorization, semantic-distance, and mixed-query assessments, the results do not clearly indicate how frequently the specific target answers remain recoverable. Additional evaluation using per-target answer recovery rates, including keyword(entity, relation) exact match (EM), would help measure whether the intended entity–relation associations are effectively suppressed.

Questions for authors

Although the theoretical motivation is reasonable, could the authors provide a direct quantitative analysis showing how the placebo correction reduces preservation-relevant changes introduced by the deletion correction? For example, analyzing token-level logit residuals on target and non-target information would help establish whether the proposed method achieves the intended effect separation.

Could the authors provide quantitative results measuring how frequently the specific target answers associated with the forget entity–relation pairs remain recoverable after unlearning? Such an evaluation, especially using keyword exact match (EM) would complement the existing aggregate and semantic evaluations and provide more direct evidence of targeted knowledge suppression.

Have the authors examined how different replacement entities or placebo relations affect the resulting contrasts? In particular, an analysis of the relationship between target–placebo semantic similarity, logit residuals, and unlearning performance would help assess the robustness of the proposed construction.

6 comments
AuthorsOpenReview ↗

Response to Weaknesses

W1: Direct Analysis of Collateral Effects during Unlearning

Thank you for the suggestion. We conduct a matched-forgetting token-level logit analysis to isolate the effect of placebo compensation from differences in forgetting strength.

Using the same deletion assistant and β−=2.1\beta_-=2.1, deletion-only (β+=0\beta_+=0) achieves MEM. =0.894=0.894, while placebo compensation with β+=0.25\beta_+=0.25 achieves MEM. =0.883=0.883. This operating point is used only for this controlled analysis; the default PLACEBU configuration remains β−=β+=2\beta_-=\beta_+=2.

At identical teacher-forced answer-token positions, we measure logit changes on 400 target C11C_{11} examples and 800 matched non-target C10/C00C_{10}/C_{00} controls.

CorrectionMEM. ↑Target-answer Δlog⁡p\Delta\log p ↓Non-target KL ↓Non-target residual RMS ↓Non-target answer Δlog⁡p\Delta\log p ↑
Deletion only0.894-22.7057.2054.385-5.324
Deletion + placebo compensation0.883-19.9945.6803.832-4.158

At closely matched forgetting strength, placebo compensation reduces non-target KL by 21.2% and non-target residual RMS by 12.6%. KL decreases for all 800 controls, and residual RMS decreases for 799/800 controls. The average non-target answer log-probability loss is also reduced from −5.324-5.324 to −4.158-4.158.

These results directly show that deletion strongly changes target-answer logits but also affects non-target information. Placebo compensation removes a substantial portion of these non-target changes while maintaining nearly the same target-forgetting strength. We will add this analysis to the revised paper.


W2: Direct Evaluation of Targeted Knowledge Suppression

Thank you for raising this point. We additionally evaluate each target entity–relation association using keyword exact match (EM) in the generated responses.

TOFU provides an original QA pair and a semantically equivalent paraphrased pair for each target fact. We identify answer-specific content words shared by both reference answers, remove stopwords and words appearing in either question, and rank the remaining keywords by IDF. Keyword EM@kk counts an answer as recovered only when the generated response contains all top-kk target keywords exactly. Lower values indicate stronger targeted suppression.

ModelKeyword EM@1 ↓Keyword EM@2 ↓Keyword EM@3 ↓
PLACEBU10.53% (36/342)3.07% (8/261)1.70% (3/176)
Retrained reference12.57% (43/342)1.92% (5/261)1.14% (2/176)

Under Keyword EM@2, PLACEBU recovers both target-answer keywords in only 8 of 261 eligible queries (3.07%), compared with 5 of 261 (1.92%) for the retrained reference. The difference is 1.15 percentage points, with an author-cluster bootstrap 95% confidence interval of [−1.58,4.18][-1.58, 4.18], which includes zero.

These results directly show that the targeted answers are rarely recoverable after unlearning and that PLACEBU’s recovery rate is statistically comparable to retraining without the forget data. We will add this evaluation alongside the existing aggregate and semantic evaluations.

AuthorsOpenReview ↗

Responses to Questions 1-3

Q1. Does placebo correction reduce preservation-relevant changes introduced by deletion?

Yes. As detailed in our response to W1, we conduct a matched-forgetting analysis at identical teacher-forced answer-token positions. At nearly the same MEM. level, placebo compensation reduces non-target KL by 21.2% and non-target residual RMS by 12.6%. It reduces KL for all 800 non-target controls and residual RMS for 799/800 controls.

These results provide direct token-level evidence that placebo compensation reduces preservation-relevant changes introduced by deletion while retaining nearly the same target-forgetting strength.


Q2. How frequently do target entity–relation answers remain recoverable?

As detailed in our response to W2, we measure per-target recoverability using Keyword EM@1, EM@2, and EM@3. Under Keyword EM@2, PLACEBU recovers both target-answer keywords in only 8 of 261 eligible queries (3.07%), compared with 5 of 261 (1.92%) for the retrained reference.

The difference is not statistically significant, indicating that PLACEBU suppresses the targeted entity–relation associations to a level comparable to retraining without the forget data.


Q3. How sensitive is PLACEBU to different replacement entities and placebo relations?

We examine this robustness using two independently generated quartet constructions on the complete TOFU Forget10 split, which contains 20 fictional authors and 400 forget examples.

For each author, we construct a unified profile and fact ledger from the corresponding 20 QA pairs. Using two different random seeds, we then independently generate the replacement-author profiles, counterfactual target facts, and placebo assignments. The original C11C_{11} cell remains fixed, while C01C_{01}, C10C_{10}, and C00C_{00} are regenerated and audited using the same construction and acceptance criteria. Both variants are trained and evaluated under identical settings.

Forget10 quartet constructionAGG. ↑MEM. ↑UTIL. ↑
Seed 17010.650.670.63
Seed 17020.660.680.65

Despite using different replacement entities, counterfactual facts, and placebo relations, the two constructions differ by only approximately 0.01 in AGG., MEM., and UTIL. This indicates that PLACEBU is robust to the particular quartet realization when the same semantic-matching and auditing criteria are applied.

This experiment evaluates construction-level robustness rather than an exhaustive sweep over target–placebo semantic similarity. Because our acceptance criteria restrict the constructions to semantically valid contrasts, the observed similarity range is limited. We have clarified this point and included the robustness results in the revised manuscript.

I appreciate the authors' efforts in conducting additional experiments and addressing my comments; however, some of my original concerns remain unresolved.

W1. Insufficient Direct Analysis of Collateral Effects

Although the additional experiments demonstrate reduced logit shifts on constructed placebo controls, it remains unclear whether these improvements extend to the preservation of actual retain knowledge. Furthermore, the analysis uses inference coefficients that differ substantially from the default configuration.

W2. Insufficient Direct Evaluation of Targeted Knowledge Suppression

While the keyword-based EM evaluation partially addresses this concern, it may fail to capture semantic recovery through paraphrasing or alternative expressions. Moreover, the absence of a statistically significant difference does not establish forgetting performance equivalent to full retraining.

Q3. Insufficient Sensitivity Analysis of Placebo Contrasts

Although the additional experiments demonstrate stability across different construction seeds and generators, they do not directly examine how target–placebo semantic similarity affects the magnitude and direction of logit residuals or the resulting unlearning performance, as originally requested.

As these concerns remain unresolved, I maintain my original score.

Reviewer_8ghSOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies unlearning without access to retain datasets. The authors argue that recent methods usually depend on retain datasets to preserve model utility after unlearning, but such datasets are often unavailable. Localized-based methods eliminate the need for retain datasets, but cannot well separate forget-specific and retain-specific effects in the latent space. The authors propose PLACEBU, a retain-free unlearning method, to address this problem.

Critical strengths
  1. Clear motivation for retain-free unlearning.
  2. The experiments cover four models and two benchmarks.
Critical weaknesses
  1. The central claim on the effectiveness of matched target-placebo examples is not supported by any evidence. It is also unclear how to construct these examples for general facts with unclear definitions of entities and relations. The paper only provides definitions of entities and relations for synthetic profiles in TOFU, while none for news articles or books in MUSE. So the effectiveness and general construction recipe of matched target-placebo examples are questionable.
  2. The retain-free claim is overstated. The author states that PLACEBU achieves unlearning without access to retain examples, but PLACEBU uses these examples to tune its hyperparameters β+,β−,κ+,κ−\beta^+, \beta^-, \kappa^+, \kappa^- (Appendix C). Further, unlearning performance is also sensitive to the choice of β+,β−\beta^+, \beta^- (Table 22).
  3. The theory of this paper is based on the existence of "admissible decompositions", but its definition is not given. Hence, Proposition 1 cannot say how much ambiguity actually exists.
  4. Δi\Delta_i and ϵctrl,i\epsilon_{\text{ctrl}, i} are not empirically quantifed. It remains unclear whether the condition ϵctrl,i<Deltai/2\epsilon_{\text{ctrl}, i} < Delta_i/2 holds in practice. This matters because this is the only theoretical result that guarantees PLACEBU improves over localization-based methods.
  5. Lemma 1 assumes β+=beta−=1\beta^+ = beta^- = 1, but experiments use β+,beta−>1\beta^+, beta^- > 1. Since \epsilon_\text{ctrl} increases as \beta^+, beta^-$ increase (Eqn. 58), the bound becomes loose.
  6. The authors make a wrong impression of PLACEBU's efficiency. The reported efficiency covers only assistant training, yet none for target-placebo construction (C10,C01,C00C_{10}, C_{01}, C_{00}) and inference. These two are expensive. Target-placebo construction requires an LLM for generation and regeneration if rejected. PLACEBU's unlearned models require four forward passes per prompt at inference time. Yet, an end-to-end complexity analysis is not provided.
  7. PLACEBU archives forgetting only at inference time and leaves the model's parameters unchanged. Hence, PLACEBU appears to suppress rather than remove knowledge from LLMs. This contradicts the motivations for selected training data removal in the introduction.

** Post-rebuttal updates**

I raised the score to 2.

Questions for authors
  1. By contrasting samples that differ in specific entities or relations, PLACEBU may rely on surface keywords to achieve unlearning. How robust is it against benign modifications of retain queries [1]?
  2. How robust is PLACEBU against adversarial attacks: GCG with adversarial suffixes, fine-tuning, orthogonalization, logit lens, and pruning [2]?
  3. Can you compare PLACEBU against all baselines on MUSE?
  4. Which LLM is used to generate C10,C01,C00C_{10}, C_{01}, C_{00}? How sensitive are the results to the choice of LLMs?
12 comments
AuthorsOpenReview ↗

Response to Weakness 1

Thank you for your insightful comments. We have carefully addressed them and uploaded our manuscript accordingly (updated text in red).

W1. Evidence and construction of matched target–placebo controls

Thank you for raising the concern about the details of MUSE construction. We address this point in two parts: (1) whether matching matters and (2) how the four cells are constructed for general passages. In addition, we have updated the manuscript and added the complete construction algorithm, cell-specific audit criteria, and full News and Books quartet examples to detail this procedure.

(1) Whether matching matters. Appendix F.2 Table 10 provides a controlled comparison in which all variants use the same number of control examples, training steps, and inference-tuning budget. It is also attached below:

Control constructionAGG. ↑MEM. ↑UTIL. ↑
Random retain control0.480.520.45
Semantic-nearest retain control0.560.500.63
Matched placebo control0.650.680.63

At the same utility of 0.630.63, the matched placebo improves MEM. from 0.500.50 to 0.680.68 over the semantic-nearest control. It also outperforms the random retain control despite requiring no original retain examples. This shows that an arbitrary auxiliary example or semantic neighbor is insufficient; preserving the same identity intervention and matching the non-target relation structure is important.

(2) Construction for MUSE. MUSE does not require a predefined knowledge-graph ontology. We define an entity as any stable, consistently replaceable participant in a passage, including a person, organization, work, location, or named group. A relation is an atomic factual or event assertion together with its semantic roles, polarity, temporal or locational qualifiers, expected answer type, and completion scope.

The passage is first decomposed into source-grounded atomic facts. For each fact with a stable replaceable participant, the four cells are constructed as follows:

  • C11C_{11}: the original identity and original target fact;
  • C01C_{01}: the dictionary-mapped identity and a counterfactual fact preserving the same target relation and role structure;
  • C10C_{10}: the original identity under a distinct non-target placebo relation;
  • C00C_{00}: the same placebo event as C10C_{10}, with only the identities replaced using the same dictionary as C01C_{01}.

Thus, C11/C01C_{11}/C_{01} share the target relation, while C10/C00C_{10}/C_{00} share the placebo relation. Across both pairs, the original-to-replacement identity intervention is held fixed. Facts are rejected if neither a stable replacement nor a valid non-target relation can be constructed.

For example, from the MUSE-News fact

The mystery act called the ChurnUps was revealed to be Foo Fighters,

we construct the fixed mapping

ChurnUps → NightSignals; Foo Fighters → Silver Atlas.

The resulting quartet is:

CellConstructed statement
C11C_{11}The mystery act called the ChurnUps was revealed to be Foo Fighters.
C01C_{01}The mystery act called NightSignals was revealed to be Silver Atlas.
C10C_{10}The ChurnUps and Foo Fighters appeared in a backstage photograph released before the festival.
C00C_{00}NightSignals and Silver Atlas appeared in a backstage photograph released before the festival.

Here, C11C_{11} and C01C_{01} instantiate the same target relation, while C10C_{10} and C00C_{00} instantiate the same independently generated placebo event. The placebo cells neither state nor entail the target fact.

The construction is completed through a generation–audit–repair pipeline, which is the key of our proposed four-cell construction. Our deterministic checks verify identity closure and exact event equivalence between C10C_{10} and C00C_{00}; while our semantic checks verify source support for C11C_{11}, target-relation preservation in C01C_{01}, and the answerability, fluency, and target exclusion from the placebo cells C10C_{10} and C00C_{00}. Rejected candidate sentences are regenerated under the same frozen article-level dictionary and construction rules.

Using the passage-level realization, PLACEBU obtains VerbMem/KnowMem/UtilPres of 10.40/20.92/52.4310.40/20.92/52.43 on MUSE-News and 9.19/4.60/60.819.19/4.60/60.81 on MUSE-Books. See more in our updated manuscript.

AuthorsOpenReview ↗

Response to Weaknesses 2 & 3

W2. End-to-end retain-free protocol and coefficient selection

Thank you for pointing out that the sentence “retain-side scores may be used” in Appendix C makes the reported protocol ambiguous. This sentence describes an optional operating-point selection procedure, rather than the protocol used for the reported PLACEBU results.

Our PLACEBU does not require retain set. As shown in our reported experiments, original retain examples are used only after configuration selection by the benchmark evaluator to report UTIL. They are not used for quartet construction, assistant training, checkpoint selection, or inference-parameter selection.

The calibration process of β−\beta_- and β+\beta_+ also does not require retain set. It works in two stages. We first select the smallest deletion coefficient β−\beta_- that satisfies a prescribed forget-side MEM. requirement, avoiding unnecessarily strong deletion. We then initialize β+\beta_+ by matching the scales of the deletion and compensation residuals, by the equation below:

β+init=β−RMS(r−)RMS(r+).\beta_{+}^{\mathrm{init}}=\beta_{-}\frac{\mathrm{RMS}(r^{-})}{\mathrm{RMS}(r^{+})}.

This normalization is needed because equal coefficients do not imply equally strong corrections when r−r^- and r+r^+ have different magnitudes. Around this initialization, we select β+\beta_+ and the support-mask parameters by minimizing residual change on the constructed C10/C00C_{10}/C_{00} placebo controls while maintaining the prescribed MEM. requirement. This procedure uses only forget queries C11C_{11}, and forget-derived controls C01C_{01}, C10C_{10}, and C00C_{00}. The resulting coefficient and mask tuple are fixed before retain-side evaluation and shared across all requests within the benchmark setting.

Table 22 of our paper shows the expected operating-point trade-off rather than dependence on retain data. Increasing β−\beta_- generally strengthens forgetting, whereas increasing β+\beta_+ restores more non-target behavior while moderately weakening forgetting. All settings reuse the same frozen assistants, and no retain examples enter either training objective or the retain-free calibration procedure.

We have removed the ambiguous sentence from Appendix C, added the above calibration procedure explicitly, and distinguished parameter selection from post-hoc utility evaluation. This is to further clarify the key contribution of this study: under the protocol used for the reported results, PLACEBU remains end-to-end retain-free.


W3. Role of the decomposition

Please note that PLACEBU does not need to find a “correct” decomposition of the hidden representation. The decomposition in the paper is used only to explain a limitation of localization.

The same localized representation may contain both information that should be removed and information that should be preserved. If two different assignments of these components produce the same observed representation, a localization-only method cannot distinguish them. If only one assignment is possible, there is no ambiguity and the lower bound becomes zero.

To avoid confusion, we will therefore replace the term “admissible decomposition” with the clearer phrase “compatible decompositions".

PLACEBU avoids choosing among these latent decompositions. It introduces a matched placebo contrast and subtracts the variation shared by the target and placebo interventions, as in the equation below:

target contrast−placebo contrast=target-specific effect+remaining control mismatch.\text{target contrast}-\text{placebo contrast}=\text{target-specific effect}+\text{remaining control mismatch}.

Thus, the decomposition motivates why localization alone can be insufficient; it is not an assumption required to run PLACEBU.

AuthorsOpenReview ↗

Response to Weaknesses 4 & 5

W4. Empirical evidence for control mismatch reduction

Δi\Delta_i and εctrl,i\varepsilon_{\mathrm{ctrl},i} are latent quantities and cannot be observed exactly. The theoretical comparison is therefore conditional rather than an unconditional claim that #\varepsilon_{\mathrm{ctrl},i}<\Delta_i/2# always holds. We have revised the manuscript to make this point explicit.

Furthermore, we test the observable implication of the matched-control argument: whether placebo compensation reduces deletion-induced changes on matched non-target information. Using the same frozen 8B Forget10 assistants, we compare deletion only with deletion plus compensation at the closest available MEM. operating points. The compensation setting is used only for this diagnostic. The results are shown below:

CorrectionMEM. ↑Non-target KL ↓Residual RMS ↓Non-target answer Δlog⁡p\Delta\log p ↑
Deletion only0.8947.2054.385-5.324
Deletion + placebo compensation0.8835.6803.832-4.158

With only a 0.0110.011 difference in MEM., placebo compensation reduces non-target KL by 21.2% and residual RMS by 12.6%. KL decreases on all 800 matched controls, and RMS decreases on 799/800 controls. The loss in non-target answer log-probability is also reduced from −5.324-5.324 to −4.158-4.158.

The above new results are not to claim the exact latent quantities, but can verify the mechanism predicted by the theory: the matched placebo contrast cancels a substantial portion of the collateral change introduced by deletion while preserving nearly the same forgetting strength.


W5. Non-unit coefficients and residual-scale calibration

Lemma 1 uses unit coefficients under the normalization that the assistant residuals already match the scales of the abstract target and placebo contrasts. In practice, residual magnitudes depend on LoRA scaling, rank, training dynamics, and the fitting quality of each assistant. Thus, even when the learned directions are correct, their raw scales may differ:

rˉi−≈s−di,rˉi+≈s+pi.\bar r_i^- \approx s_-d_i,\qquad \bar r_i^+ \approx s_+p_i.

The coefficients β−\beta_- and β+\beta_+ calibrate these implementation-dependent scales. Placing the approximation assumptions directly on the deployed residuals,

∥β−rˉi−−di∥≤ε−,iβ,∥β+rˉi+−pi∥≤ε+,iβ,\|\beta_-\bar r_i^- - d_i\| \leq \varepsilon_{-,i}^{\beta}, \qquad \|\beta_+\bar r_i^+ - p_i\| \leq \varepsilon_{+,i}^{\beta},

gives

∥β−rˉi−−β+rˉi+−τi∥≤εctrl,i+ε−,iβ+ε+,iβ.\|\beta_-\bar r_i^- - \beta_+\bar r_i^+ - \tau_i\| \leq \varepsilon_{\mathrm{ctrl},i} +\varepsilon_{-,i}^{\beta} +\varepsilon_{+,i}^{\beta}.

Therefore, non-unit coefficients do not increase the matched-control error εctrl,i\varepsilon_{\mathrm{ctrl},i}; they set the learned residuals to their corresponding theoretical contrasts.

Appendix B.5 expresses non-unit coefficients through a conservative deviation term relative to the unit-coefficient normalization. We have added the above scale-calibrated formulation and clarified that the last Equation in Appendix B.5 is a conservative bound, rather than evidence that control mismatch grows with β\beta. Table 22 empirically shows the resulting smooth forgetting–utility trade-off.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.