acceptodds
Under review as a conference paper at ICLR 2027

CURE: Counterfactual Unlearning of Large Language Models via Density Ratio Estimation

Abstract

Machine unlearning for large language models aims to remove the influence of designated training data while preserving behavior supported by the remaining data. Yet knowing what to forget does not determine what the model should predict after forgetting: the factual prediction entangles retain- and forget-supported influences, making the desired retain-only behavior unidentifiable from the original model alone. We show that this ambiguity can be resolved by introducing a controlled second observation that amplifies the forget influence. Under a retain–forget mixture model, our Bidirectional Mixture Identification result reduces recovery of the retain-only next-token distribution to estimating a context-dependent forgetting ratio. We further show that this ratio is characterized by the density ratio between the factual and forget-amplified models and can be updated autoregressively during generation. Based on these results, we propose CURE (Counterfactual Unlearning via Ratio Estimation), which uses a forget-aligned auxiliary model to reveal the direction of removal and contextual ratio estimation to determine its magnitude, while keeping the factual model fixed. We additionally analyze how model-realization, auxiliary-alignment, and ratio-estimation errors affect recovery. Across TOFU, MUSE-News, and RealToxicityPrompts, CURE achieves favorable forgetting–utility trade-offs; on TOFU-10%, it attains forget/retain ROUGE-L of 38.3/98.4, compared with 39.6/99.2 for the retain-only reference. These results support counterfactual unlearning as an identification-and-estimation problem rather than solely an objective for suppressing forget-associated behavior.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Model, Unlearning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_C4MQOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes CURE (Counterfactual Unlearning via Ratio Estimation). The key idea is to identify the ratio/weight of the forget set's contribution in the desired retain-only model. To do so, the paper uses an auxiliary model that amplifies the forget influence. CURE estimates the forget-source weight by tracking how much more likely each prefix is under the auxiliary model than under the factual model and updates it autoagressively. CURE then use the estimated weight to recover the retain-only next-token distribution. CURE is empirically evaluated on TOFU, MUSE, and RealToxicityPrompts.

Critical strengths
  1. The paper raises an important perspective. To recover the model or distribution after unlearning, we may first need to identify the weight/ratio the forget set play in generating the output sequences.
Critical weaknesses
  1. The paper is not clearly presented and it is difficult to follow. This makes it hard to accurately assess the contributions and soundness of the paper.

    • Although Secs 2-5 of the paper describe what is done, they consistently do not describe why it is done and do not provide intuition, a top down overview, or an explanation of each theorem's significance for the readers. In my opinion, significant rewriting may be needed.
    • How is pRp_R different from qRq_R? Is the goal of unlearning to recover pRp_R?
    • Sec 2.1 suggests that the weight rtr_t must be known to uniquely recover pRp_R.
    • Sec 2.2 suggests a way to set rtr_t but does not justify why. How is α(x)\alpha(x) chosen or known? (Eq 16 is still unclear) The subsection then abruptly introduces an auxiliary model that amplifies the forget distribution.
    • In Sec 3, the concept "decoder state" r^t\hat{r}_t is suddenly introduced. Why is the recovery operator Rκ\mathcal{R}_\kappa and FF introduced and as defined in equation 8? Rκ\mathcal{R}_\kappa seems to be designed to recover pRp_R based on eq (6) and (7). FF is not even used in the main paper.
    • It is hard to understand and comment on the remaining theoretical results due to the lack of clarity. The theoretical results seem to bound the error in estimating qRq_R and rtr_t
    • A better explained algorithm pseudocode in the main paper would be useful
  2. The experiment results can be more complete. See questions below.

Questions for authors
  1. Address the questions and concerns in weakness 1. Additionally, explain the intuition/significance of each theorem and design choice.
  2. Discuss the computation cost/complexity of CURE as compared to other unlearning methods.
  3. Beyond the usual unlearning metrics in Table 1 and 2, is it possible to report how CURE’s next-token predictions match those of the retrain model or report the approximation error? This would better relate the experiment results to the theory section.
  4. Are the experiments per section of the table repeated for different seeds? No uncertainty estimate is reported.
  5. if the claim is a better forget-utility tradeoff, it would be better to plot forget against retain ROUGE in a graph to compare methods across iterations as well.
  6. On what dataset or forget-retain split is CURE expected to perform best or worst? When is CURE preferred over other methods?

Other suggestions

  • Bold and underline the best and second best result in each column of the experiment results
4 comments
AuthorsOpenReview ↗

Response to Reviewer C4MQ (1/4)

Q1(a). The practical target and the ideal distribution

The practical unlearning target is qR=A(Dr)q_R=A(D_r), the predictor trained on retained data under a specified training procedure, including its initialization and randomness. The distribution pRp_R is the ideal retain-source component in the mixture model. Introducing pRp_R makes it possible to study which predictive observations identify the retained contribution and to derive a corresponding recovery rule.

Theorem 3.1 connects this identification result to the practical target. Equation (11) bounds the discrepancy between the repaired prediction and qRq_R using model-realization, auxiliary-alignment and state-estimation errors, repair distortion, and ϵR=∥qR−pR∥1\epsilon_R=\lVert q_R-p_R\rVert_1. This gives a quantitative connection between the identifiable retain-source component and the practical retraining target. The response to Question 3 evaluates next-token agreement with qRq_R directly.

Q1(b). Why the contextual weight is needed

The contextual weight resolves an ambiguity in the observed predictions. Consider a two-token illustration with pD=(0.6,0.4)p_D=(0.6,0.4) and pF=(0.8,0.2)p_F=(0.8,0.2). Under pD=(1−r)pR+rpFp_D=(1-r)p_R+rp_F, the choice r=1/4r=1/4 yields pR=(8/15,7/15)p_R=(8/15,7/15), whereas r=1/2r=1/2 yields pR=(0.4,0.6)p_R=(0.4,0.6). Both explanations produce the same observed pDp_D and pFp_F.

Theorem 2.1 identifies the missing information: specifying a feasible weight determines which retain prediction is intended. This motivates the next step of the method—obtaining additional information through a controlled auxiliary observation and calibrated prefix ratios. The required weight concerns the current prefix, so it generally differs from the fraction of training examples designated for deletion.

Q1(c). Prompt posteriors and initialization

In the calibrated construction, α(x)\alpha(x) is derived by processing the prompt from a known BOS source prior α0\alpha_0. With known amplification parameter κ\kappa and the assumed joint observation laws,

1+κα(x)=(1+κα0)pA(x)pD(x).1+\kappa\alpha(x)=(1+\kappa\alpha_0)\frac{p_A(x)}{p_D(x)}.

Prompt prefill determines the source weight for the first response token by incorporating the prompt's source evidence. The same likelihood-ratio update then adapts the correction strength as the response develops. For ideal response likelihood ratios and the unprojected state, Equation (16) quantifies how the initial error is multiplied by the response-prefix density ratio.

The reference initialization a0=β/(κλ)=0.5a_0=\beta/(\kappa\lambda)=0.5 is specified by the training weights; under BOS calibration it represents α0\alpha_0. The archived decoder uses a differently parameterized state initialized at zero and omits prompt prefill.

AuthorsOpenReview ↗

Response to Reviewer C4MQ (2/4)

Q1(d). Why introduce controlled amplification

Controlled amplification provides a second observation of the same source components with a different forget contribution. For rt<1r_t<1, combining that observation with the factual mixture eliminates the unknown forget conditional and gives

pR=pD−1+κrtκ(1−rt)(pA−pD).p_R=p_D-\frac{1+\kappa r_t}{\kappa(1-r_t)}(p_A-p_D).

This equation explains the design: pA−pDp_A-p_D reveals the direction introduced by amplification, while the contextual coefficient determines how far to reverse that movement. Calibrated prefix density ratios supply the coefficient. The auxiliary and ratio estimator therefore provide complementary information for identifying the retain prediction.

Theorem 5.1 connects this construction to auxiliary training: it characterizes the data conditional under which the reverse-KL objective realizes the amplified observation. The result gives a precise training condition for producing the second observation used by the recovery rule.

Q1(e). Recovery and state-update operators

Each generation step has two tasks. The recovery operator RκR_\kappa first uses the current source weight to correct the next-token prediction. After a token is selected, the prefix changes, so the next correction must incorporate the new source evidence. The update operator performs this second task:

\qquad \ell_t=\frac{p_A(y_t\mid h_t)}{p_D(y_t\mid h_t)}.$$ A token more likely under the amplified observation increases the unprojected state; a token less likely under it decreases the state. The recursion consequently changes the correction strength along the observed prefix. The reference decoder evaluates these operations with neural likelihood ratios, and Theorem 3.1 quantifies how state-estimation error affects the recovered prediction. ### Q1(f). The theorem-to-algorithm connection The results form the following progression: | Question | Role of the result | |---|---| | What information is missing? | Theorem 2.1 relates the unidentified contextual weight to ambiguity in the retain prediction. | | What additional observations suffice? | Theorem 4.1 gives recovery from controlled amplification, calibrated BOS information and prefix ratios. | | What must auxiliary training realize? | Theorem 5.1 characterizes the required auxiliary-data conditional. | | How does approximation affect the practical target? | Theorem 3.1 separates the errors contributing to disagreement with retraining. | The strict reference decoder follows this sequence: initialize at BOS; update through prompt tokens; evaluate both next-token distributions; apply recovery and probability repair; select a token; and update that beam's state. The strict upper margin keeps the inverse finite, while repair handles negative coordinates caused by approximation. On the same 20 QA examples per split, the historical decoder and complete strict reference yield: | Decoder | Forget ROUGE | Retain ROUGE | |---|---:|---:| | Historical CURE | 32.133 | 99.000 | | Complete strict reference | 83.989 | 77.950 | This shared-subset comparison directly tests how decoder choice affects the forgetting–retention balance. The historical decoder achieves lower forget ROUGE and higher retain ROUGE, linking the high-retention result to a concrete decoding configuration.
AuthorsOpenReview ↗

Response to Reviewer C4MQ (3/4)

Q2. Preparation, inference and implementation cost

CURE keeps the factual model fixed, trains an auxiliary, and performs additional model evaluation during decoding. Table 8 reports the following TOFU-10% system measurements:

MethodPreparation, minInference, minPeak inference memory, GB
DPO883515
NPO543515
GA623515
ULD367221
CURE75054

The measurements locate the preparation–inference trade-off: CURE takes 7 minutes of preparation and 50 minutes of inference, compared with 54 and 35 minutes for NPO. With beam width BB, response length TT and vocabulary size VV, probability correction adds O(TBV)O(TBV) arithmetic. The archived implementation also performs full-vocabulary sorting, adding O(TBVlog⁡V)O(TBV\log V), and recomputes complete prefixes without incremental KV reuse. Model-forward cost additionally depends on prefix length and model size.

Reusing an identical auxiliary forward result reduces two auxiliary calls per native step to one. In the reuse experiment, generation time falls from 4.2899 to 3.5731 seconds, a 16.7% reduction, with all 66 checked token sequences unchanged. The sequence comparison ties the speed gain to preserved outputs, showing that this optimization removes redundant computation.

Q3. Direct next-token comparison with retraining

We compare factual and CURE next-token distributions with the retain-only retrained predictor at common prefixes. The evaluation uses 100 QA examples per split and measures TV after probability repair; lower values indicate closer agreement:

Splitn QATV factualTV CURE
Forget1000.3949150.841886
Retain1000.0106670.006092

On Retain, CURE reduces TV by 0.004575, approximately 42.9% relative to the factual model. This directly answers whether the retained predictions become closer to the retrained predictor: the correction improves next-token agreement on Retain. On Forget, TV increases by 0.446971, placing the corrected prediction farther from that reference. The comparison ties empirical evaluation to the practical recovery target defined in the theory.

Q4. Repeated training and uncertainty

To test whether the retention result persists across auxiliary-training seeds, the added study repeats complete auxiliary training with seeds 0, 101 and 202, using 20 epochs and 8,000 updates per run. Each trained auxiliary is evaluated on the same 25 QA examples per split:

Training seedForget ROUGERetain ROUGE
038.537100.000
10132.894100.000
20233.54298.857
Mean ± sample SD34.99 ± 3.0999.62 ± 0.66

Across the three trained auxiliaries, retained ROUGE averages 99.62 with a sample standard deviation of 0.66, while forget ROUGE averages 34.99 with a sample standard deviation of 3.09. Keeping the evaluation questions fixed makes this a direct check of sensitivity to auxiliary training. High retained ROUGE recurs in all three runs, addressing the concern that retention depends on one favorable auxiliary-training seed. Appendix D separately reports 100 controlled synthetic trials per sample size.

AuthorsOpenReview ↗

Response to Reviewer C4MQ (4/4)

Q5. Checkpoint trajectories

The checkpoint comparison follows CURE auxiliary epochs 2, 5, 10 and 20 and full-model NPO steps 1, 4, 16 and 32 on the same ten forget and ten retain QA examples:

Method / checkpointQA per splitForget ROUGERetain ROUGE
CURE consistency .5 / epoch21038.301989.1935
CURE consistency .5 / epoch51029.411392.5329
CURE consistency .5 / epoch101033.47199.2593
CURE consistency .5 / epoch201041.4255100
NPO full / step11099.3849100
NPO full / step41042.99861.861
NPO full / step16104.44634.4302
NPO full / step321000

Forget and retain ROUGE at saved checkpoints, using the same 10 QA examples per split. Labels indicate auxiliary-training epochs for CURE and optimizer steps for NPO.

CURE's retained ROUGE increases from 89.19 at epoch 2 to 100 at epoch 20, while forget ROUGE varies non-monotonically. NPO moves toward stronger suppression together with reduced retained ROUGE. Among these saved checkpoints, CURE at epoch 20 and NPO at step 4 have comparable Forget ROUGE (41.43 versus 43.00), while Retain ROUGE is 100.00 versus 61.86. This comparison demonstrates CURE's retention advantage at this tested forgetting level. Following both scores across saved checkpoints also makes the effect of operating-point selection explicit.

Q6. Dataset differences and deployment choices

The analysis predicts favorable recovery conditions when the auxiliary is well aligned with the controlled observation, initialization is calibrated, and relevant prefixes have adequate coverage and retain-source support. The inverse becomes more sensitive as the source posterior approaches one or the amplification parameter approaches zero. This explains why forget-dominated prefixes are difficult: the inverse amplifies model and state-estimation errors as the retain-source contribution shrinks.

The TOFU results show how behavior varies with the deletion split:

Forget splitCURE forget / retain ROUGE-LRetain-only forget / retain ROUGE-L
1%46.7 / 99.542.7 / 98.9
5%38.1 / 93.439.2 / 98.9
10%38.3 / 98.439.6 / 99.2

Among these runs, the 5% split has the largest retained-score deficit, 5.5 percentage points; the 10% run is close to the reference on both generation scores. On MUSE, retained utility is 69.53 versus 69.05 for retraining on Books, and 52.1 versus 55.7 on News. These comparisons answer the suitability question in terms of retained behavior: the current configuration achieves near-reference utility on Books, with larger retention gaps on News and TOFU-5%.

CURE is useful when an application needs to preserve the factual checkpoint and can accommodate an auxiliary model and a custom decoder. Its intervention operates during decoding, with the factual weights unchanged. Inference-memory and latency requirements can be assessed using the resource measurements in our response to Question 2.

Other suggestion. Table markings

For MUSE privacy deviation, values closer to zero rank higher; the other columns follow the metric directions stated in their captions. The retain-only row provides the target behavior against which the unlearning methods are compared. These criteria give a consistent basis for marking the best and second-best entries across columns.

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

This paper frames LLM unlearning as a statistical problem of identifying the retain-source distribution within a mixture. It assumes a context-dependent mixture of retain-source and forget-source distributions, and derives a relation between the contextual source weight and the retained conditional distribution. The paper proposes to construct forget-related data and train an auxiliary model to approximate a controlled amplification of the forget source. Under its stated mixture and calibration assumptions, it uses prefix density ratio tracking from the beginning of the sequence to estimate the source weight. It then combines the predictions from the frozen original model and trained auxiliary model using the source-weight estimate to recover a corrected distribution. It corrects next-token probabilities during decoding using beam search with width 4. The paper analyzes the calibration conditions and approximation errors, and reports evaluations on TOFU, MUSE, and RealToxicityPrompts.

Critical strengths
  1. The paper proposes a recovery mechanism with useful conditional analysis. Theorem 4.1 derives the recovery formula and prefix-based source-weight update under the stated mixture and calibration assumptions. Theorem 3.1 quantifies the sensitivity of recovery to model mismatch, auxiliary-alignment error, and estimation error. These are useful conditional analyses of the particular construction.

  2. The reported evaluation on TOFU and MUSE shows promising results for the trade-off between forgetting and retention. On TOFU-10%, CURE’s forget/retain ROUGE-L is 38.3/98.4, close to the retain-only reference’s 39.6/99.2. On MUSE Books, knowledge/verbatim memorization decreases from the original model’s 45.88/99.70 to 12.01/9.64, while retained utility is 69.53, compared with 69.05 for Retrain.

Critical weaknesses
  1. The paper proposes a statistical framework relevant to unlearning. However, it remains unclear why this framework adequately models the LLM setting. There is a gap that occurs when it moves from partitioning training data into retain and forget sets to assuming a mixture of retain-source and forget-source prediction distributions. A partition of the training data does not automatically imply that a trained neural network’s predictions admit the required decomposition. The paper introduces ideal distributions for this purpose, but their correspondence to the original and retrained LLMs needs stronger justification.

  2. The proposed method relies on a forget-aligned auxiliary, but Section 5 and Appendix E.1 describe its examples as supplied or generated and filtered without a reproducible per-benchmark recipe, including generation prompts, filtering rules, and proposal/acceptance counts. This makes it difficult to assess how much the results depend on additional data or curation and whether baselines receive comparable information. The detailed preparation-and-inference cost comparison is limited to TOFU-10%, a small synthetic unlearning task, and does not clarify whether data construction is included. Although MUSE provides useful real-data evidence, the paper does not establish how auxiliary quality, required data, and total cost change with the scope of knowledge to be removed. Also, RMU is compared only on TOFU, is absent from the MUSE comparison, and is not evaluated against CURE on WMDP, where it was introduced. A reproducible construction protocol, raw-data versus curated-data controls at comparable budgets, and a quality and cost study on WMDP or a comparable broader task would strengthen the evidence for practical applicability.

  3. The motivation is not clearly positioned within the established discussion of the forgetting–retention trade-off. Although the introduction acknowledges the difficulty of removing forget-data influence while preserving retained behavior, the subsequent emphasis on distribution recovery does not sufficiently explain how this formulation advances existing approaches to that problem. Retain-only recovery specifies a more precise target, yet retraining references, retention objectives, and explicit prediction targets already play important roles in LLM unlearning. Although Appendix A discusses several related methods, their connection to the proposed recovery target and relevant post-training work needs clearer explanation. The paper would benefit from clarifying which limitations of existing approaches its recovery formulation addresses, what contribution comes specifically from controlled amplification and recursive ratio estimation, and which empirical evidence demonstrates the resulting benefit. This would better establish the significance of the method beyond its reported forgetting–utility results.

  4. The ablations do not sufficiently isolate CURE’s core components. Although the paper reports several sensitivity studies, the comparison of adaptive ratio estimation with alternative controls is limited to two examples. A full-test-set comparison against a tuned fixed-coefficient correction, using the same auxiliary and decoding budget, would clarify the benefit of recursive estimation. Separate ablations of BOS prompt prefill and reverse-KL auxiliary regularization would further help identify which components drive the reported gains.

  5. The performance comparison is insufficiently controlled and documented. Table 8 reports 7 minutes of preparation for CURE versus 54 for NPO. CURE uses an 8-layer LoRA auxiliary, whereas the baseline protocol describes full-parameter target-model training. Also, the TOFU-10% GA/NPO configurations have zero model utility, so the timings do not establish efficiency at comparable forgetting–utility operating points. Cross-method search budgets are not established as matched: CURE uses beam width 4, while baseline settings are incompletely specified, affecting interpretation of both generation scores and inference costs. Training-budget descriptions and timing boundaries, including data construction and reference caching, need clarification.

Questions for authors
  1. Could the authors clarify the connection between the proposed theory and the practical implementation? Which of the mixture and calibration assumptions are supported empirically for the trained LLMs? In particular, how does the initialization a0=0.5a_0=0.5 relate to the BOS source prior, and what empirical evidence supports identifying it with that prior? What supports the correspondence between the ideal retain-source distribution pRp_R and the retrained predictor qRq_R?

  2. Regarding auxiliary-data construction and evaluation coverage (Weakness 2), could the authors clarify how the auxiliary examples are obtained for each benchmark, including any generation/filtering procedure, data size, and use of retained examples? Are data construction costs included in Table 8's preparation time? Also, could the authors explain why RMU was omitted from the MUSE comparison and whether CURE was evaluated on WMDP? If such results and auxiliary-construction costs are available, they would help assess whether the method generalizes to broader knowledge-removal settings.

  3. Could the authors provide further details on the training and generation configurations used for the main results and Table 8 timings? In particular, how do the epoch and update counts in E.2–E.3 correspond, and which beam widths and sampling settings were used for each method and the retrained reference? If GA/NPO configurations preserving useful retained performance were evaluated, how do their quality and costs compare?

  4. Would the authors be able to provide a full-test-set comparison between CURE's recursive estimator and a tuned fixed-coefficient correction, using the same auxiliary, probability repair, and decoding budget? Such a comparison would help assess the contribution of adaptive estimation beyond auxiliary prediction subtraction and search, complementing the reported alternative-control comparison on two examples.

5 comments
AuthorsOpenReview ↗

Response to Reviewer wFXG (1/4)

Q1(a). Predictive mixtures and the retraining target

The practical target is the retain-only retrained predictor, qR=A(Dr)q_R=A(D_r). Under the predictive-mixture model, CURE identifies the ideal retain component pRp_R from the factual and controlled amplified observation laws, using calibrated prefix ratios and the required BOS prior. Theorem 3.1 connects this construction to qRq_R through a bound containing model-realization, auxiliary-alignment, state-estimation, and repair errors, together with ϵR=∥qR−pR∥1\epsilon_R=\lVert q_R-p_R\rVert_1. The bound identifies the quantities controlling approximation to the retrained predictor. Appendix D tests the construction under known source laws, and the neural measurements below directly compare CURE with the retrained reference.

The common-prefix comparison directly tests the practical recovery target by comparing next-token probabilities at identical prefixes. It measures mean TV distance to qRq_R after probability repair; lower is better.

SplitQA itemsFactual TVCURE TV
Forget1000.3949150.841886
Retain1000.0106670.006092

CURE reduces Retain TV by approximately 42.9%, supporting the retention objective through closer next-token agreement with the retrained predictor. On Forget, TV increases from 0.394915 to 0.841886, a change of 0.446971.

Q1(b). Initialization and state interpretation

The BOS prior supplies the starting information for contextual ratio estimation. Under calibrated joint observation laws, Proposition B.1 propagates a known BOS source posterior through prompt likelihood ratios to obtain the prompt-specific posterior; the same recursion then processes response tokens. The reference initialization a0=β/(κλ)=0.5a_0=\beta/(\kappa\lambda)=0.5 follows from the chosen loss weights. Its posterior interpretation requires prior compatibility, a0=α0a_0=\alpha_0, calibrated joint observation laws, and clipping that leaves the ideal recursion unchanged. Prompt prefill converts the starting prior into a correction strength informed by the actual input, explaining how the construction obtains context-specific control. Appendix C.4 quantifies how initialization error propagates through the subsequent likelihood-ratio updates.

AuthorsOpenReview ↗

Response to Reviewer wFXG (2/4)

Q2(a). Auxiliary-data construction

The documented TOFU-10% auxiliary uses the following recipe:

ComponentConfiguration
InitializationFirst eight layers, embeddings, normalization, and output head of the factual checkpoint
Forget examples400 supplied QA pairs
Trainable parametersAll eight-layer auxiliary parameters; 20 epochs and 8,000 updates
Retain consistencyFrozen initial-eight-layer oracle; KL(oracle8∥qA)\mathrm{KL}(\mathrm{oracle8}\Vert q_A), weight 0.5, temperature 2

The forget examples train the auxiliary observation, while the retain-consistency branch encourages preservation of predictions supported by retained examples. The factual checkpoint remains fixed throughout this adaptation. Listing both branches makes the additional training information explicit: the auxiliary uses forget examples together with a retention constraint. The training-data paraphrasing experiment below examines how the resulting system responds to changes in the wording of the forget examples.

Q2(b). Robustness to training-data paraphrasing and removal scope

CURE maintains a Retain ROUGE-L score of 92.156% when 50% of the forget training data are semantically rewritten. This directly addresses dependence on the original forget-data wording: high retained-task performance persists after substantial changes to that wording, providing evidence of retention robustness under the tested condition.

Forget training data paraphrasedForget ROUGE-L (%)Retain ROUGE-L (%)
0%27.483999.8519
50%37.829392.156

Relative to the original-data condition, Retain ROUGE-L decreases from 99.8519% to 92.156%, a change of −7.6959 percentage points, while Forget ROUGE-L increases from 27.4839% to 37.8293%, a change of +10.3454 percentage points. Thus, substantial paraphrasing preserves a high level of retained-task performance, with reduced forgetting strength and a measurable retention cost.

Appendix E.9 evaluates incremental adaptation from TOFU-1% to 5% and then 10%. Starting from the existing auxiliary, one adaptation epoch with retained-example replay takes 58 and 108 seconds, respectively, excluding initial training. Retain ROUGE-L is 80.3 and 85.2 at these stages, compared with 93.4 and 98.4 in the corresponding full-training study. This sequence addresses how CURE handles expanding deletion requests: an existing auxiliary can be updated rapidly, with the measured retention trade-off shown by the full-training comparison.

Q2(c). RMU comparison and task coverage

The additional RMU evaluation directly addresses the missing representation-based baseline on the real-text MUSE News and Books tasks. It uses the official domain targets, 512-token windows, seed 42, batch size 4, learning rate 5×10−55\times10^{-5}, retention weight 100, and steering coefficient 20, updating the down projections in layers 5–7. News/Books use 508/553 updates and 100 records per final metric. CURE entries below are the published results, retained without reevaluation.

DomainMethodKnow ↓Verb ↓Priv → 0Util ↑
NewsCURE10.40011.400−8.61052.100
NewsRMU64.10255.036−99.81155.738
BooksCURE12.0109.640−26.50069.530
BooksRMU42.1453.335−5.08162.804

CURE has lower knowledge memorization in both domains and higher Books utility, supporting its ability to reduce memorized content while retaining useful predictions on real text. RMU provides higher News utility and better Books verbatim/privacy results. The comparison makes these practical choices visible across the two domains. CURE has not been evaluated on WMDP.

AuthorsOpenReview ↗

Response to Reviewer wFXG (3/4)

Weakness 3. Contribution to the forgetting–retention problem

CURE addresses a specific ambiguity within the established forgetting–retention problem: when retained and forgotten data both support a continuation, the factual prediction alone does not determine how much of that prediction should remain after removal. Retain-only retraining supplies the evaluation target, while CURE identifies the additional information needed to construct a prediction toward that target. Under the shared-source mixture assumptions, controlled amplification supplies a second observation; calibrated prefix density ratios and the required BOS prior determine the contextual correction magnitude. This yields a probability-space inverse with coefficient γκ(rt)=(1+κrt)/(κ(1−rt))\gamma_\kappa(r_t)=(1+\kappa r_t)/(\kappa(1-r_t)). In comparison, the fixed logit contrast in Appendix A uses a constant exponent on the auxiliary probabilities. CURE therefore connects the removal direction and magnitude to an identification problem, with an error analysis quantifying the effects of auxiliary alignment, initialization, state estimation, and probability repair.

Q4. Recursive estimation versus fixed logit correction

The fixed-logit comparison tests whether a global correction strength provides the desired forgetting–retention balance. The three coefficients expose how changing this strength affects both tasks:

MethodFixed logit coefficientForget ROUGE-L (%)Retain ROUGE-L (%)
Original CURE—34.603199.2
Fixed logit correction0.2557.894497.1613
Fixed logit correction0.523.634396.3613
Fixed logit correction0.919.716286.7149

CURE reaches 99.2% Retain ROUGE-L, the highest value in this comparison. Against the fixed coefficient 0.25, it also lowers Forget ROUGE-L from 57.8944% to 34.6031%. Coefficients 0.5 and 0.9 yield lower forget scores, with retain scores of 96.3613% and 86.7149%, respectively. Increasing the coefficient from 0.5 to 0.9 gains 3.9181 points of additional forget-overlap reduction at a cost of 9.6464 retain points. This answers why correction magnitude matters: stronger global subtraction can also remove retained behavior. CURE supplies a prefix-dependent alternative and, in this comparison, combines greater suppression and retention than the tested coefficient 0.25 while preserving more retained behavior than 0.5 or 0.9.

Weakness 4. Prompt prefill and auxiliary regularization

The prefill comparison changes this component within the complete strict reference, evaluated on 20 QA items per split. ROUGE is reported on a 0–100 scale:

DecoderForget ROUGERetain ROUGE
Strict reference with prefill83.98977.950
Strict reference without prefill78.22878.564
Historical decoder32.13399.000

Within the strict reference, enabling prefill increases Forget ROUGE by 5.761 points and decreases Retain ROUGE by 0.614 points. Thus, prefill weakens forgetting and slightly reduces retained performance in this configuration. This on/off comparison directly answers whether prefill improves these generation scores. The historical decoder reaches 99.0 Retain ROUGE with 32.133 Forget ROUGE.

Retain consistency substantially improves retained-task performance at a fixed 20-epoch/8,000-update training budget, evaluated on 25 QA items per split. ROUGE uses the same 0–100 scale.

Consistency weightForget ROUGERetain ROUGE
038.89551.895
0.538.537100.000

Adding KL(oracle8∥qA)\mathrm{KL}(\mathrm{oracle8}\Vert q_A) with weight 0.5 raises Retain ROUGE from 51.895 to 100.000, a gain of 48.11 percentage points. Forget ROUGE changes from 38.895 to 38.537. The large retain improvement alongside a similar forget score explains the role of the regularizer: it protects retained behavior as the auxiliary is trained on forget examples.

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

This paper studies LLM unlearning as recovery of retain-only predictions, rather than simply suppression of forget-associated outputs. Under a retain–forget mixture model, it introduces a controlled auxiliary observation that amplifies the forget component and uses prefix density ratios to identify context-dependent correction coefficients. The resulting method, CURE, combines a frozen factual model, a forget-aligned auxiliary model, and recursive probability-space correction. The paper provides conditional identification and approximation results and evaluates the method on TOFU, MUSE, and RealToxicityPrompts. The reported results show promising forgetting–utility trade-offs, although the connection between the ideal recovery framework and the implemented LLM system remains insufficiently established.

Critical strengths
  1. The predictive target is well motivated. Distinguishing retain-only recovery from indiscriminate suppression is important: information associated with forgotten examples may also be supported by retained data. Framing this ambiguity as an identification problem provides a useful perspective on unlearning.

  2. The framework makes its assumptions and error sources explicit. The analysis separates model realization, auxiliary alignment, ratio estimation, and probability repair. The discussion of instability near a source posterior of one, and the characterization of auxiliary calibration in Theorem 5.1, help clarify when the proposed recovery interpretation is applicable.

  3. The empirical trade-offs are encouraging. On TOFU-10%, CURE reports forget/retain ROUGE-L of 38.3/98.4, compared with 39.6/99.2 for the retain-only reference. On MUSE-Books, retained utility is 69.53 versus 69.05 for retraining. The auxiliary-sensitivity, computational-cost, and repeated-extraction analyses provide useful additional context. The manuscript also appropriately acknowledges that these behavioral measurements do not establish distributional equivalence.

Critical weaknesses
  1. Recovery of the ideal source distribution is not yet connected convincingly to the retraining target. The intended target is qR=A(Dr)q_R=A(D_r), while the identification results recover an ideal distribution pRp_R. Equation (11) includes their discrepancy ϵR\epsilon_R, but the experiments do not establish that this term is small. A partition of the training data does not itself imply the corresponding mixture structure for neural predictors. Consequently, the reported generation metrics support behavioral utility more directly than counterfactual distribution recovery. This distinction is acknowledged in the paper, but remains central to assessing its contribution (Sections 2–3 and 6.1).

  2. The practical auxiliary construction does not establish the required calibration. Theorem 5.1 characterizes an auxiliary-data conditional involving the desired amplified distribution. However, the experimental construction does not demonstrate that its forget-related data approximates this conditional. Appendix C.2 provides a revealing counterexample showing that ordinary forget-data NLL with reverse-KL regularization need not realize controlled amplification. Similarly, the implemented initialization a0=0.5a_0=0.5 is not established as the BOS source posterior. These gaps limit the interpretation of the recursive decoder state as an identified forgetting ratio.

  3. The executed decoder is not fully specified or aligned with the analyzed algorithm. Algorithm 1 requires a strict upper margin below one, whereas Appendix E.1 reports clipping to [0,1][0,1]. The recovery formula is singular at the upper endpoint, and the manuscript leaves the exact endpoint treatment and numerical smoothing value unspecified. Appendix F.2 also describes a repair map different from Equation (18). This does not establish that the actual runs failed numerically, but it prevents precise reproduction and leaves the applicability of the strict-margin analysis unresolved.

  4. The experiments do not isolate the benefit of recursive density-ratio control. The adaptive-versus-prescribed-schedule comparison contains two examples rather than a test-set ablation, and Appendix A acknowledges differences in generation rules. Comparisons using the same auxiliary model and decoding budget are needed to distinguish the proposed mechanism from fixed probability subtraction or fixed logit contrast. Aggregate ROUGE-L and a high KS-test p-value are insufficient for this attribution; common-prefix distribution measurements would be substantially more informative.

Questions for authors
  1. Using the existing factual, auxiliary, and retain-only models, can the authors report common-prefix distribution comparisons? In particular, does qA−qDq_A-q_D approximately align with qD−qRq_D-q_R, and does CURE reduce TV or KL distance to qRq_R before and after probability repair?

  2. How does the actual auxiliary-data construction approximate the conditional required by Theorem 5.1? What evidence supports interpreting a0=0.5a_0=0.5, and the subsequent recursive state, as the relevant source posterior?

  3. Which exact decoder produced the main tables? Please specify the upper-endpoint rule, numerical constants, and repair map, and clarify whether the strict-margin reference implementation reproduces the reported behavior.

  4. On one representative existing benchmark, how does recursive control compare with fixed probability-space correction and fixed logit contrast when the auxiliary model and decoding budget are held constant? Such a comparison would directly clarify the contribution of the proposed adaptive mechanism.

4 comments
AuthorsOpenReview ↗

Response to Reviewer zjDh (1/4)

Q1. Direction, magnitude and agreement with retraining

CURE separates the direction of removal from its magnitude. Under the shared-source mixture and controlled amplification assumptions, Eq. (43) gives, at a common prefix with 0<rt<10<r_t<1,

pA−pD=κ(1−rt)1+κrt(pD−pR).p_A-p_D=\frac{\kappa(1-r_t)}{1+\kappa r_t}(p_D-p_R).

The positive coefficient means that amplification moves along the displacement away from the ideal retain prediction. Reversing this shift therefore supplies the recovery direction. This is the purpose of the controlled second observation: it exposes a direction that the factual predictor alone cannot determine. For trained models, the corresponding empirical prediction is alignment between qA−qDq_A-q_D and qD−qRq_D-q_R, where qRq_R is the retain-only retrained predictor.

The magnitude is identified from the joint prefix ratio, given the BOS source prior α0\alpha_0 and calibrated joint observation laws:

\frac{p_A(x,y_{<t})}{p_D(x,y_{<t})},\qquad p_R=p_D-\frac{1+\kappa r_t}{\kappa(1-r_t)}(p_A-p_D).$$ Equations (14)–(15) therefore determine a contextual correction strength from prefix evidence. A strict margin below one controls the inverse's sensitivity. Theorem 3.1 connects approximate recovery to the practical target $q_R$ through model-realization, auxiliary-alignment, state-estimation and repair errors, together with $\lVert q_R-p_R\rVert_1$. Appendix D tests whether the identified direction and magnitude recover the known retain law, using the terminal-projected estimator under synthetic finite-sequence laws. The exact-law control has average recovery TV of $1.33\times10^{-15}$. Across 100 trials per sample size, increasing sequences per observation law from 200 to 20,000 reduces ratio MSE from 0.12717 to 0.00662 and repaired recovery TV from 0.28488 to 0.08067; factual-to-retain TV is 0.1. The exact-law result verifies the recovery calculation numerically, and the finite-sample results connect improved ratio estimation to improved recovery. At the largest reported sample size, estimation yields a closer retain prediction than the factual distribution. To test agreement with the practical retraining target, the same-prefix neural comparison evaluates factual and repaired CURE predictions against $q_R$, using 100 QA examples per split. Lower TV indicates closer agreement. | Split | QA examples | Factual TV | CURE TV | |---|---:|---:|---:| | Forget | 100 | 0.394915 | 0.841886 | | Retain | 100 | 0.010667 | 0.006092 | CURE reduces Retain TV by 0.004575, or approximately 42.9%. This directly addresses target agreement on retained content: the correction brings its next-token distribution closer to the retain-only retrained predictor. Forget TV increases by 0.446971.
AuthorsOpenReview ↗

Response to Reviewer zjDh (2/4)

Q2(a). Auxiliary construction and calibration

The auxiliary is designed to provide the controlled second observation needed to identify the removal direction. Theorem 5.1 makes this requirement precise: for the stated reverse-KL objective, the amplified conditional is the unique population optimum exactly when the auxiliary-data conditional satisfies Eq. (22), subject to its feasibility condition. Proposition C.1 extends this connection to approximate calibration, separating optimization error from a stationarity residual that captures auxiliary-data and reference-model mismatch. Appendix C.2 explains why this condition matters: in its counterexample, ordinary forget-data NLL with reverse KL cannot realize the specified amplified target for any positive loss weights. Together, these results give auxiliary construction an explicit identification criterion and a bound for approximate training.

The recorded TOFU auxiliary uses 400 supplied QA examples and 20 epochs/8,000 updates. Its consistency term is KL from a frozen copy of the initial eight-layer model to the auxiliary, with weight 0.5 and temperature 2. The controlled ablation, evaluated on 25 QA examples per split, shows its contribution to retained behavior:

Consistency weightForget ROUGERetain ROUGE
038.89551.895
0.538.537100.000

Adding the consistency term improves Retain ROUGE by 48.11 percentage points while changing Forget ROUGE by only 0.358 points. This comparison tests whether the auxiliary-training constraint helps preserve retained behavior, and the retention gain explains its practical role in the auxiliary construction.

Q2(b). BOS initialization and the recursive state

The role of BOS tracking is to obtain the context-dependent correction magnitude from prefix evidence. Under the joint observation laws in Eq. (25), Theorem 4.1 and Proposition B.1 give

1+κrt=(1+κα0)pA(x,y<t)pD(x,y<t).1+\kappa r_t=(1+\kappa\alpha_0)\frac{p_A(x,y_{<t})}{p_D(x,y_{<t})}.

Starting from the source prior α0\alpha_0, processing prompt tokens therefore yields the prompt-conditioned posterior and continues through generation. The implementation rule a0=β/(κλ)=0.5a_0=\beta/(\kappa\lambda)=0.5 specifies a decoder initialization. Its posterior interpretation requires a0=α0a_0=\alpha_0, calibrated joint laws, and clipping that preserves the ideal recursion. Equation (16) quantifies how initialization error propagates when the response likelihood ratios are ideal. These conditions give the recursive weight its statistical meaning through the source prior and the accumulated prefix evidence.

The inspected historical decoder initializes a differently parameterized state at b0=0b_0=0, omits prompt prefill, and uses a piecewise update. The reference-decoder comparison isolates initialization and prefill changes on 20 QA examples per split:

Reference configurationForget ROUGERetain ROUGE
Prefill, initial state 0.583.98977.950
Without prefill78.22878.564
Prefill, initial state 083.98977.950

Within this strict reference, enabling prefill weakens forgetting and slightly reduces retained performance: Forget ROUGE rises by 5.761 points and Retain ROUGE falls by 0.614 points. The two tested initial values yield identical reported scores when prefill is enabled. These controls directly answer the implementation-sensitivity question by showing the observed effect of each execution choice.

AuthorsOpenReview ↗

Response to Reviewer zjDh (3/4)

Q3. Executed decoder and strict-reference comparison

All numerical safeguards use the same value, 0.000001 (1e-6). This includes the probability floor after the inner correction, the denominator floor in the outer correction, the offset used in probability repair, the denominator offset in the likelihood ratio, and the offset added before taking log probabilities. The protected denominator remains positive when the recursive state reaches its endpoint.

The implementation first forms an inner correction from the factual and auxiliary predictions, applies the probability floor, and normalizes. It then applies a state-dependent outer correction. Probability repair subtracts the smallest coordinate, adds the common offset, and renormalizes. The state starts at zero without prompt prefill, follows the implemented piecewise likelihood-ratio update after each selected token, and is clipped to the interval from zero to one. Beam selection uses mean log probabilities; correction logits and the outer states and coefficients are computed in float32. This specifies the endpoint handling and repair order behind the decoder's output.

The complete strict reference uses an amplification factor of 1, an initial state of 0.5, prompt prefill, a state cap of 0.9, a single recovery step, the probability repair in Eq. (18), accumulated log scores, and stopping when a generated token is an EOS token. On 20 QA examples per split:

DecoderForget ROUGERetain ROUGE
Historical decoder32.13399.000
Complete strict reference83.98977.950

The complete reference does not reproduce the historical trade-off on this subset, directly answering the question about replacing the full execution procedure. A separate cap-only sensitivity check in Table 16 reports forget/retain ROUGE-L of 42.9/97.2 at cap 0.9 versus 38.3/98.4 at cap 1. This second comparison measures how forgetting and retention change when the state margin alone is adjusted within that configuration.

AuthorsOpenReview ↗

Response to Reviewer zjDh (4/4)

Q4. Recursive control and fixed corrections

CURE's recursive state adjusts the correction strength using prefix evidence. Theorem 4.1 identifies this contextual quantity under calibrated observation laws. The fixed-logit ablation compares CURE with three constant correction strengths:

MethodFixed logit coefficientForget ROUGE-L (%)Retain ROUGE-L (%)
Original CURE—34.603199.2
Fixed logit correction0.2557.894497.1613
Fixed logit correction0.523.634396.3613
Fixed logit correction0.919.716286.7149

CURE has the highest retain score among these settings. Relative to the fixed coefficient 0.25, it lowers Forget ROUGE-L by 23.2913 percentage points and raises Retain ROUGE-L by 2.0387 points. Across these fixed-logit settings, larger coefficients are accompanied by progressively lower forget overlap and lower retention. For example, moving from 0.5 to 0.9 lowers Forget ROUGE-L by another 3.9181 points while reducing Retain ROUGE-L by 9.6464 points. The sweep demonstrates the importance of correction magnitude and identifies CURE's high-retention operating point. Its contextual correction rule addresses precisely this magnitude-selection problem.

A separate controlled neural experiment supplies direct evidence about state estimation because its source law and initial prior are known. At κ=1\kappa=1, evaluations on training contexts averaged over three seeds at 1,024 updates show that dynamic estimation reduces branch-position state MAE from 0.159360 to 0.101772, demonstrating more accurate tracking of the contextual weight at these positions. TV to the retain-only predictor at branch positions decreases from 0.879818 for fixed-state correction to 0.864525 for dynamic correction, while factual TV is 0.204558. Across all positions, dynamic TV is 0.338507 versus 0.306338 for fixed-state correction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.