acceptodds
Under review as a conference paper at ICLR 2027

CINDER: Twin-Referenced Training for Relearning-Resistant Unlearning

Abstract

LLM unlearning methods that pass their own evaluations are routinely undone by a few steps of fine-tuning on part of the forget set. We argue the right reference for resistance is not "unrecoverable" but a retain-only twin — a model trained without the forget data — and train against it directly. CINDER simulates the attacker's own attack class inside training (full-parameter fine-tuning on a random half of the forget facts, first-order meta-gradient) and penalizes recovery of the held-out half through a softplus penalty anchored at the twin's post-attack floor. We evaluate with the hypotheses the right way round: method-level non-inferiority to the twin at a pre-specified margin — the spread between two independent retrains — across corpora, with every axis required. On sixteen corpus seeds at 0.5B, CINDER's post-attack recovery exceeds the twin's by 0.08 nats (one-sided 95% bound 0.21, margin 0.36), 1.17 nats below NPO (95% CI 0.94–1.39); on sixteen fresh corpora generated after the analysis was fixed, +0.16 nats (bound 0.29), 1.16 below NPO. It is non-inferior on six of eight axes; on the circuit criterion and utility non-inferiority is not shown. A frozen, preregistered eight-axis protocol, used as a screen, agrees: relearning is not rejected on 15 of 16 seeds (16 of 16 fresh) against at most 4 for five baseline families, and no axis on 8 of 16 (3 of 16 fresh, C1 rejecting eleven) against at most 2. Ablations over eight to sixteen seeds show that the floor, the full-parameter inner attack and the fact-level split are each needed, that a floor taken from the pretrained base — no twin at training time — works as well as the twin's, and that adding our floor to MUDMAN brings it within the margin. At 1.5B and 3B relearning is not rejected on any of twelve seeds, but utility is paid for. We report what this does not show: a pass of a finite battery bounds nothing, the objective overshoots its floor, and a twin, though not needed for training at 0.5B, is still needed to evaluate.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
machine unlearning, relearning attacks, retain-only twin, non-inferiority testing, meta-learning, LLM safety

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_mqAmOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes CINDER, a meta-unlearning method designed to resist subsequent relearning. During training, CINDER simulates a full-parameter fine-tuning attack on one half of the forget facts and penalizes recovery on the held-out half relative to a retain-only twin. Experiments on synthetic entity-QA corpora show that its post-attack recovery is non-inferior to the twin and substantially lower than NPO across both headline and fresh seeds. Additional evaluations cover multiple recovery attacks, objective ablations, model scales, and limited experiments on TOFU and MUSE, although non-inferiority is not established for circuit removal or retained utility.

Critical strengths
  • CINDER introduces a clear and meaningful objective for relearning resistance by directly simulating a full-parameter attack and comparing recovery with a retain-only model that never observed the forget data.

  • The evaluation is unusually transparent and thorough, with multiple corpus seeds, a separate set of fresh seeds, extensive ablations, held-out attacks, multi-axis testing, and detailed reporting of negative results and failed configurations.

Critical weaknesses

i) Although the twin is presented as a reference for calibrated forgetting, CINDER does not consistently stop once the forget-set likelihood reaches the twin level. At 0.5B, the final forget log-probability is on average 1.08 nats below the twin, and this gap grows to 6.7 nats in the 5% forgetting experiment. The ablation results further show that preventing this overshoot substantially reduces relearning resistance: using a fixed coefficient or stopping at the twin boundary decreases the number of passed relearning tests from 15/16 to 10/16 and the number of certified cases from 8/16 to 1/16. This suggests that the strongest results may come from aggressive suppression rather than the proposed twin-matching principle.

Such suppression does not necessarily mean that the underlying information has been removed. Prior work shows that apparently forgotten knowledge can return after benign relearning [a] and may remain recoverable from internal representations even when output-level accuracy is low [b]. The paper should separate the benefit of twin guidance from the benefit of pushing the forget examples farther below their original likelihood, and should report the resulting effect on semantically related knowledge and downstream utility. This is particularly important because the larger-model experiments already show unstable utility and retention, including rejection of all six utility tests at 1.5B.

ii) Most of the evidence comes from Qwen2.5-0.5B on a synthetic entity-QA dataset in which the target facts are introduced during supervised fine-tuning. Because the retain-only twin shares the same pretrained base, this setting mainly tests whether CINDER can remove newly injected facts; it does not show whether the method can remove knowledge acquired during pretraining. This distinction matters because unlearning methods evaluated on fine-tuning data can appear successful while pretrained information remains recoverable [c]. The TOFU and MUSE experiments provide some additional evidence, but each uses only one seed and covers a limited subset of the available evaluation dimensions. On TOFU, the mean truth ratio also falls to 25.7 compared with 1.04 for the twin, again indicating substantial over-forgetting rather than close behavioral matching.

Scalability is also not demonstrated experimentally. Both the inner and outer optimization loops update full model parameters, and the estimated memory requirement increases from 12.9 GB at 0.5B to 199.4 GB at 7B. The main 0.5B experiments are also substantially slower than NPO and, on the reported testbed, even slower than retraining the retain-only twin. The 1.5B and 3B results show unstable utility, while the 7B setting is supported only by a memory estimate rather than an actual unlearning experiment. The paper should validate the method on larger pretrained models, naturally occurring knowledge, and multiple random seeds, while reporting end-to-end time and memory against exact retraining and other relearning-resistant baselines.

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

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

[c] Do Unlearning Methods Remove Information from Language Model Weights? arXiv 2024.

Questions for authors
  1. How much of CINDER’s relearning resistance comes from matching the retain-only twin, and how much comes from suppressing the forget-set likelihood below the twin? Can you report results under an enforced stopping rule that prevents this overshoot?

  2. Does CINDER remain effective when the forgotten knowledge was acquired during pretraining rather than introduced through supervised fine-tuning? Can you provide multi-seed results on TOFU, MUSE, or another realistic benchmark?

  3. What are the measured runtime, peak memory, and utility results on larger models such as 3B or 7B, and how do these costs compare with exact retraining and other relearning-resistant methods?

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

This paper proposes CINDER, a machine unlearning method designed to improve resistance to relearning attacks. The method uses a retain-only twin, trained without the forget data, as a reference for the desired forgetting behavior. During training, CINDER randomly splits the forget facts into two subsets, simulates a relearning attack by fine-tuning the model on one subset, and penalizes knowledge recovery on the held-out subset using a twin-referenced objective. The authors evaluate CINDER against existing unlearning methods across multiple attack settings, model sizes, and random seeds. Experimental results show improved resistance to relearning attacks, although the method still suffers from over-suppression and utility degradation, particularly on larger models.

Critical strengths
  1. Meaningful research problem. The paper addresses an important problem in machine unlearning by investigating whether forgotten knowledge can be recovered through relearning attacks.
Critical weaknesses
  1. Poor presentation and unclear motivation. The paper does not clearly explain the background and limitations of existing unlearning methods, nor does it provide a smooth transition to the problem of relearning resistance. In particular, the motivation for studying relearning attacks and using a retain-only twin as a reference is not sufficiently developed. The organization is also confusing. Section 3, titled Preliminaries, mainly discusses the evaluation protocol, including attack settings and statistical tests, rather than introducing the background concepts needed to understand the method. This makes the paper difficult to follow and obscures the main research contribution.
  2. Insufficient explanation of the proposed method. The method section is too brief and lacks sufficient details to explain the proposed approach. For example, the authors simply split the forget set into two halves for attack simulation and evaluation, but do not clearly explain why this design is appropriate or how it helps achieve relearning resistance. The motivation behind several key design choices is also insufficiently discussed. Overall, Section 4 reads more like a brief summary of the method rather than a complete methodological description. At the very least, an overview figure illustrating the workflow and the relationships between different components would help readers better understand the proposed approach.
Questions for authors
  1. What is the rationale for randomly splitting the forget facts into two equal halves? Why is a 50/50 split appropriate, and how sensitive is the proposed method to different splitting ratios?
  2. Could the authors provide a more detailed explanation of how the simulated relearning attack and the twin-referenced penalty work together to achieve relearning resistance?
  3. Although CINDER shows improved relearning resistance, the results also reveal over-suppression and utility degradation. How can the authors establish that CINDER achieves better unlearning rather than simply suppressing knowledge more aggressively? Would the improvements still hold when comparing methods at similar levels of retain utility?
Reviewer_uF7GOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes CINDER, a method that removes facts from a language model so that fine-tuning cannot easily bring them back. During training, CINDER attacks a copy of the model with one half of the forget facts. Then it penalizes the recovery of the other half (the held-out half). The penalty uses a retain-only twin as its reference, which is a model that never saw the forget data. The authors test CINDER against the twin with a non-inferiority test on 16 synthetic data sets with Qwen2.5-0.5B. Then they repeat the test on 16 new data sets.

Critical strengths
    1. Good objective. CINDER does not try to get zero recovery. It tries to get the same recovery as a model that never saw the forget data. Thus, a method that damages the model to remove the facts does not get a reward.
    1. Clear reporting. The paper gives its limits in detail. It also gives ablation results, withdrawn claims and one configuration error.
Critical weaknesses
    1. Limited experimental evidence. The main results use one 0.5B model and synthetic knowledge that fine-tuning added. The TOFU and MUSE-News tests use only one seed each and only some of the evaluation checks (lines 336–342). Most importantly, the paper does not show the method on knowledge from pretraining. At 1.5B, all six seeds fail the utility check (lines 428–429). This causes concern about scalability and practical use.
    1. Limited attack types and possible Goodhart effects. CINDER trains against full-parameter fine-tuning on forget facts, and the evaluation mainly uses the same attack class. The held-out attacks change only the number of epochs, the learning rate, the data split and one LoRA setting. They do not test attack types that are fundamentally different (lines 324–326). Tests with adaptive attacks, much longer fine-tuning or fine-tuning on unrelated data would better show that the resistance applies to attacks other than the training attack.
    1. Presentation and reproducibility. The abstract is dense and has many numbers in each sentence (lines 010–032). The paper uses some terms before it defines them, or never defines them: "UMA" (line 142), "λ2/λ3" (line 216, defined at line 995), "ε-rehearsal" (line 266) and "non-naive seeds" (line 251). The eight checks come from a certificate that is not published (line 463). Table 4 has four empty "Per-method summary" headings (lines 681–688). The paper gives the method only as a numbered list (lines 172–179), and a figure or an algorithm box would help.
Questions for authors
    1. The paper says that a floor from the pretrained base gives similar results at 0.5B (lines 447–449). What does the retain-only twin add during training?
    1. Can the authors test more types of fine-tuning attacks? For example: longer attacks, adaptive attacks and fine-tuning on unrelated data.
    1. Can the authors show unlearning of knowledge from pretraining? If they cannot, can they explain how the method applies to this case
Reviewer_sh1pOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes CINDER, a first-order meta-unlearning method designed to resist fine-tuning relearning attacks up to the level of a "retain-only twin" (an exact retrain baseline fine-tuned without the forget set). On every second retain step, CINDER splits the forget facts in half, simulates K=3K = 3 full-parameter AdamW fine-tuning steps on one half, and applies a normalized softplus penalty to the held-out half's log-probability relative to a post-attack floor ff, updating the model via FOMAML. Using an 8-axis evaluation harness from concurrent unpublished work on a synthetic 80-entity QA corpus at Qwen2.5-0.5B (across 16 headline and 16 fresh seeds), CINDER achieves non-inferiority to the twin on the relearning axis (+0.08 and +0.16 nats vs. a 0.36-nat margin), while failing non-inferiority on circuit attribution (C1) and utility. Scaling to 1.5B and 3B preserves relearning resistance on the synthetic corpus but incurs severe utility degradation and requires 28 bytes/parameter in memory.

Critical strengths

None

Critical weaknesses
  1. Practical contradiction of the problem setup, broken "floor" mechanism, and severe scaling/utility failures:

    • Contradictory setup & limited novelty: Section 1 motivates unlearning for models "whose training cannot be repeated," yet headline CINDER ("Twin-Referenced Training") explicitly trains a retain-only twin (i.e., the exact Retrain baseline) first just to compute the scalar floor ff—taking 7x longer than exact retraining itself (153–174 s vs. 25 s, Section 5.6). While a 0.5B-only ablation (floor from base, Table 3) estimates ff from the pre-finetuning base model M0M _ 0 (halving full-screen passes from 8/16 to 4/16 and worsening utility), both variants require a clean pre-exposure checkpoint M0M _ 0 and 6 full passes over the retain set RR. This leaves no viable use case: for pretraining knowledge (where training truly cannot be repeated), no clean M0M _ 0 exists so CINDER cannot be used (Section 6); for fine-tuning knowledge where both M0M _ 0 and RR exist, fine-tuning M0M _ 0 on RR directly is faster, uses half the memory, and yields an exact clean model. Meanwhile, first-order meta-learning against relearning attacks was already introduced by TAR [1] (omitted from experiments) and MUDMAN [2] (which already passes relearning on 14/16 seeds vs. CINDER's 15/16, Table 2).
    • Self-defeating floor normalization & failure to scale: Although pitched as anchoring unlearning at ff rather than over-suppressing, normalizing P(θ)P(\theta) by ∣P(θ)∣|P(\theta)| cancels the softplus exponential decay below ff (Appendix B.12), creating a constant push that overshoots the twin (by up to 6.7 nats). When the coefficient is frozen or replaced by a true ReLU hinge so optimization actually stops at ff, relearning resistance collapses (1/16 and 0/7 passes, Table 3). Requiring fp32 full-parameter AdamW in both loops costs 28 bytes/parameter (~200 GB at 7B, which is not run). Utility fails on 6/6 seeds at 1.5B, suffers catastrophic retain collapse (EM = 0.085, Table 20) under fragile non-monotonic LR tuning at 3B, fails utility and C1 non-inferiority even at 0.5B (passing all axes on only 3/16 fresh seeds), and is tested on only 1 seed of tiny TOFU/MUSE subsets (failing TOFU Forget Quality at p=0.045p = 0.045 with mean truth ratio 25.7 vs. 1.04).
  2. Poor presentation, inaccurate technical statements, and non-self-contained dependency on unpublished work:

    • Unclear method and missing definitions: The paper contains zero displayed equations or algorithm blocks. The proposed method is compressed into 15 inline lines (Section 4.1) with inconsistent notation (writing P(θ)P(\theta) on line 170 vs. θ′\theta' on line 177), undefined dataset structure ("forget texts grouped by fact" used before explaining subjects/facts/paraphrases), ff used in Step 3 before being defined, and basic details like training duration hidden on p. 23. Furthermore, evaluation relies entirely on an unpublished concurrent submission (Anonymous, 2026), leaving core metrics (UMA, C1, C2, and the $\varepsilon$-rehearsal) undefined in Section 3.
    • Inaccurate claims and editing defects: Section 1 makes false statements—claiming unlearning methods are "evaluated by the loss they were trained to minimize" (benchmarks evaluate output metrics like QA accuracy or Truth Ratio, not training losses, and output-metric fragility is already well established) and claiming quantization recovers knowledge "without touching a weight" (weight quantization directly modifies every weight). The prose reads like a cryptic internal lab log filled with undefined shorthand ($\lambda$2/$\lambda$3 on p. 5, defined on p. 19; self-contradictory text on p. 9 stating "The twin must exist at training time ... not for training"), broken table references (Table T2b on p. 6, Tables T6 and T6b on p. 9, Tables T1–T8 on p. 10, Table T2d on p. 23), and four repeated empty summary headers on p. 13.

References:

  • [1] Tamirisa et al. (2025). Tamper-resistant safeguards for open-weight LLMs. ICLR.
  • [2] Sondej et al. (2025). Robust LLM unlearning with MUDMAN: Meta-unlearning with disruption masking and normalization. arXiv:2506.12484.
Questions for authors

Clarifications:

  1. Practical applicability: If pretraining knowledge cannot be unlearned because no clean base M0M _ 0 exists (Section 6), and fine-tuning setups that retain M0M _ 0 and RR can simply train the exact retain-only twin in 1/7th the time and half the memory of CINDER (Section 5.6), in what concrete deployment scenario is CINDER preferable to exact retraining?
  2. Comparison to TAR and floor behavior: Why was TAR [1] not evaluated alongside MUDMAN, and does any variant whose gradient genuinely vanishes at the floor ff (without canceling the softplus decay via ∣P(θ)∣|P(\theta)|) maintain relearning resistance?
  3. Scaling & real benchmarks: Why does utility fail on 6/6 seeds at 1.5B and collapse on seed 45 at 3B, and how does CINDER perform on full standard TOFU and MUSE splits beyond single-seed 100–160 QA subsets at 0.5B?

Suggestions:

  • Make Sections 3 and 4 self-contained with explicit equations, pseudocode, and metric definitions independent of (Anonymous, 2026), correct the inaccurate statements in Section 1 (regarding loss evaluation and quantization), and fix the broken table references and formatting errors in the Appendix.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.