Machine unlearning means removing the influence of specific training examples from a trained model without retraining it from scratch. Its usual failure is collapse. The optimizer keeps pushing the "forget" examples after they have already been forgotten, and in doing so damages general ability. Two hand-set knobs decide how far it pushes: how many update steps to run, and which examples to keep updating. This proposal replaces both knobs with a statistical test.
After each round of updates, we compare every forget example against a reference pool of similar examples that the model never trained on. An example whose score is still unusually low compared with that pool still looks "memorized," so it keeps receiving updates. An example that now looks like the unseen pool is left alone. An example pushed far past the unseen pool, toward strangely high loss, is pulled back. Because the reference pool is re-scored by the current model at every round, the decision boundary follows the model as its weights drift across a sequence of deletion requests. Fixed thresholds calibrated once at the start cannot do this.
The move is borrowed from out-of-distribution detection, the problem of flagging inputs unlike the training data. A Provable Decision Rule for Out-of-Distribution Detection turns detector scores into p-values against a calibration set and applies a Benjamini–Hochberg-style procedure. That procedure bounds the false discovery rate (FDR), meaning the expected fraction of flagged items that were wrongly flagged. We apply the same logic to unlearning. The "detections" are forget examples still distinguishable from unseen data. The guarantee bounds the fraction of updates spent on examples that were already forgotten. Those wasted updates are exactly what drives collapse.
Core Claim
Gold-standard retraining does not drive forget examples to maximal loss. It leaves them looking like data the model never saw. PeriUn found that retrained models agree with the original model on high-confidence "core" forget examples and differ mainly on "peripheral" ones near decision boundaries. Unlearning the core examples is therefore unnecessary damage.
PeriUn implements this insight with a fixed bottom-$\alpha$ confidence quantile. That cutoff is wrong whenever a request's mix of core and peripheral examples differs from what $\alpha$ assumes. The OOD literature documents the same failure for fixed thresholds when outlier ratios shift (C2AE; Taming False Positives in OOD Detection with Human Feedback).
Our claim is that a per-example test against an unseen reference pool, re-run every round, does three things together:
- It recovers PeriUn's core/peripheral split without any $\alpha$.
- It stops unlearning automatically with no epoch count.
- It prevents the cumulative over-forgetting that FIT identifies as the driver of collapse under sequential requests.
Mechanism
Let $s(x;\theta)$ be a membership score of example $x$ under current weights $\theta$. Concretely, $s$ is the loss under $\theta$ minus the loss under the pre-fine-tuning base model; subtracting the base loss removes per-example difficulty. Let $R=\{r_1,\dots,r_n\}$ be the reference pool of held-out examples drawn from the same distribution as the forget set but never trained on.
For forget example $x_i$, the lower-tail conformal p-value is
$$p_i^{\text{mem}} = \frac{1 + \#\{j : s(r_j;\theta) \le s(x_i;\theta)\}}{n+1}.$$
It tests the null hypothesis "$x_i$ behaves like unseen data" against the alternative "$x_i$ is still memorized." An upper-tail p-value $p_i^{\text{over}}$ is defined symmetrically to detect over-forgetting.
Each round, Benjamini–Hochberg at level $q$ is applied to $\{p_i^{\text{mem}}\}$ to select an active set that receives the unlearning loss. A second Benjamini–Hochberg pass on $\{p_i^{\text{over}}\}$ selects a repair set. For these examples, a small term pulls the score back toward the reference median, rather than toward the uniform target that AdaProb imposes. Every other forget example contributes no gradient that round. Unlearning stops when the active set is empty. The base optimizer (gradient difference or NPO, each paired with a retain-set penalty) is unchanged; our rule only decides which forget examples it sees and when it stops.
Under exchangeability between a forgotten example and the reference pool, the procedure inherits the FDR bound of the multiple-testing literature. Here that bound says that at most a $q$ fraction of active examples, in expectation, were already indistinguishable from unseen data.
Two technical points need care. First, TOFU questions cluster by fictitious author, so the test is run at the author level, aggregating the scores of each author's questions. This avoids treating correlated questions as independent hypotheses. Second, a gradient step on the forget set can shift reference scores too. This is intended: re-scoring $R$ under the current $\theta$ every round is how the boundary tracks drift. Knowledge is Never Enough and Noisy Test-Time Adaptation in Vision-Language Models show that frozen boundaries go stale exactly in this setting, where model weights keep changing.
Relation to Prior Work
Work that motivates the proposal shows uncontrolled pushing causes collapse. On TOFU, none of gradient ascent, gradient difference, KL minimization, or preference optimization achieved effective unlearning, and 5% and 10% forget sets made models unusable. FIT shows that collapse compounds over sequential requests. FIT counters it with redundancy filtering and an importance score that routes requests between algorithms, but its loss-difference and importance thresholds are fixed once and never re-calibrated as the model drifts.
Work that supplies the method is the multiple-testing rule of Ma et al.. We reuse its calibration-then-BH structure, but apply it to forget examples against an unseen pool rather than to test inputs against in-distribution data.
The closest existing combination of adaptive gating and unlearning is Dynamic SAE Guardrails. It decides at inference time, using an empirically tuned token-ratio threshold, whether to clamp sparse-autoencoder features. Our rule differs on three counts: it governs training-time weight updates rather than inference-time clamping, it acts per forget example rather than per incoming prompt, and its threshold comes from a calibration guarantee rather than tuning.
Several papers select data or parameters but leave their final cutoffs heuristic:
- ALISE selects forget examples on a Pareto front.
- WAGLE selects weights by implicit-gradient attribution, and its authors say the sparsity ratio still needs manual tuning.
Our rule selects examples, not weights, so it is orthogonal to WAGLE and can be stacked on top of it.
Work that constrains the proposal comes from the safety side. An Illusion of Unlearning? shows that output-level forgetting can leave internal features intact and recoverable. RMU from WMDP deliberately scrambles internal representations for that reason. "Indistinguishable on output loss" may therefore be too weak a stopping criterion. We treat this as a test, not an assumption. We evaluate relearning attacks directly, and we run a variant whose score $s$ is a linear probe on hidden states.
The method does not naturally apply to WMDP-style hazardous knowledge, because pretraining knowledge has no "never-seen" counterpart drawn from the same distribution. We scope the claim to example-level deletion.
Research Program
We first build the reference-pool splits. For TOFU, we fine-tune Llama2-7b on the 200 authors minus a held-out block of authors, matching TOFU's own protocol otherwise. The held-out authors form $R$. Gold models retrained on retain-only data give TOFU's forget-quality reference. For PCH, we hold out a matched subset of the synthetic personal, copyright, and harmful samples in the same way. For the image setting, a held-out split of TinyImageNet and CIFAR-10 plays the role of $R$ for VGG16, matching PeriUn.
We then implement the two-sided gated procedure on top of gradient difference and NPO, plus Random Labeling for the image setting. We study the sensitivity to $q$, which replaces two knobs (epoch count, plus $\alpha$ or a sparsity fraction) with one interpretable error rate. The real question is whether results are flat across $q \in [0.05, 0.2]$.
Finally, we verify the guarantee empirically. Using the gold retrained model, we label each forget example as truly distinguishable or not, then measure the realized fraction of updates spent on already-forgotten examples against $q$.
Baselines
Naive reference points. The do-nothing point is the original fine-tuned model. The naive unlearning reference is dense gradient difference and dense NPO run for a fixed number of epochs, with the epoch count tuned on a validation forget split. This is the standard practice that TOFU reports as collapsing at 5% and 10%.
WAGLE + NPO. The strongest single-request competitor on TOFU is WAGLE with NPO, which reports unlearning efficacy 0.9599 on Llama2-7b against 0.9581 for dense NPO. We run WAGLE with its sparsity swept on validation under the same total gradient-step budget as our method. We also run WAGLE combined with our gate, because the two are complementary. Beating it means equal or higher TOFU forget quality, measured as the KS-test p-value against the retrained model, together with higher model utility at the 5% and 10% splits, without per-split tuning of anything except $q$.
FIT. The strongest sequential competitor is FIT on PCH. FIT reports the best trade-off between forget degree and retain utility, but no reported number was found on Lacuna for its PCH forget degree, retain utility, MMLU, CommonsenseQA, or GSM8K scores. We will therefore rerun FIT with its published thresholds on the same request sequence, base model, and compute. Beating it means a retain-utility and GSM8K curve that stays higher over hundreds of requests at matched forget degree.
PeriUn. In the image setting the competitor is PeriUn, which improved Tug-of-War by 22.04 points and the membership-inference gap by 8.7 points over the strongest baseline on TinyImageNet with VGG16. We run it with $\alpha$ tuned on the standard request composition and then held fixed, and compare under matched compute.
Ablations isolating the key ingredient. The first isolates batch adaptivity: replace Benjamini–Hochberg with a single per-example threshold at the reference pool's $q$-quantile. The second isolates drift tracking: calibrate $R$ once at round zero and freeze it. The third removes the repair pass.
Why not just…?
Why not just tune the number of epochs on a validation split? Epoch count is a single global stopping time, but forget examples reach the unseen level at very different rates. Core examples start there, while peripheral ones need many steps, as PeriUn's discrepancy analysis shows. Any global stopping time therefore over-pushes some examples and under-pushes others. The TOFU collapse at larger forget fractions is consistent with this. The tuned-epoch baseline above measures it directly. If tuned early stopping matches us on TOFU's fixed splits, the remaining advantage must appear on shifted compositions and in sequential streams, where a single tuned value cannot transfer.
Why not just use a calibrated per-example quantile threshold instead of FDR control? This is a real threat, and it is the first ablation. A per-example threshold re-calibrated on $R$ every round already captures the drift-tracking benefit. Benjamini–Hochberg adds value only when the fraction of still-memorized examples varies across requests, which is the regime where Ma et al. found batch-adaptive rules cut KNN false positives from 48.21% to 16.79% on CIFAR-10 versus Places365. If the two tie on our composition-shift experiment, we will report that the contribution is reference-calibrated, drift-tracking stopping, not multiple testing.
Why not just keep PeriUn's quantile and FIT's thresholds but re-tune them per deployment? Re-tuning needs a gold retrained model or a labeled validation stream for every new request mix. That is unavailable in deployment and is exactly the cost unlearning tries to avoid. Our rule needs only a held-out pool, scored by the current model. The composition-shift and sequential experiments test whether fixed-but-tuned settings actually degrade.
Why not just scramble representations like RMU so stopping does not matter? RMU targets hazardous capability rather than example-level deletion. For deletion, PeriUn's evidence suggests that pushing core examples is gratuitous damage. The relearning-attack evaluation will reveal whether our gentler stopping leaves recoverable traces. If it does, the hidden-state-probe score is the remedy we test.
Proposed Experiments
Composition shift. On TOFU and TinyImageNet, we build forget requests whose fraction of core examples ranges from 10% to 90%. On TOFU, core examples are authors whose facts the base model already predicts well. Fixed-$\alpha$ PeriUn, tuned-epoch NPO, and the fixed-quantile ablation should lose forget quality or utility as the mix moves away from their tuning point. Our method should stay flat across the mix.
Drift. On PCH, we run hundreds of sequential requests and plot retain utility, MMLU, and GSM8K against request index for our method, its frozen-calibration ablation, and FIT. The prediction is that the frozen variant and FIT degrade later in the stream, while re-calibrated gating does not.
Error-rate fidelity. For each round we compute the realized fraction of updates spent on examples the gold retrained model deems already forgotten, and compare it with $q$. If this fraction exceeds $q$ substantially, the exchangeability assumption fails in practice, and the method is reduced to a heuristic.
Robustness to relearning. Following An Illusion of Unlearning?, we fine-tune each unlearned model briefly on a small subset of the forget data and measure how quickly forget performance recovers, alongside a membership-inference attack. Matching the retrained model here, not just on output loss, is the bar for claiming that "indistinguishable from unseen" is the right stopping target.
References
- A Provable Decision Rule for Out-of-Distribution Detection
- PeriUn: Enhancing Unlearning by Selectively Forgetting Peripheral Samples
- C2AE: Class Conditioned Auto-Encoder for Open-Set Recognition.
- Taming False Positives in Out-of-Distribution Detection with Human Feedback
- FIT: Defying Catastrophic Forgetting in Continual LLM Unlearning
- AdaProb: Efficient Machine Unlearning via Adaptive Probability
- Knowledge is Never Enough: Towards Web Aided Deep Open World Recognition
- Noisy Test-Time Adaptation in Vision-Language Models
- TOFU: A Task of Fictitious Unlearning for LLMs
- SAEs Can Improve Unlearning: Dynamic Sparse Autoencoder Guardrails for Precision Unlearning in LLMs
- Alignment-aware Data Selection for Unlearning in Contrastive Vision-Language Models
- WAGLE: Strategic Weight Attribution for Effective and Modular Unlearning in Large Language Models
- An Illusion of Unlearning? Assessing Machine Unlearning Through Internal Representations
- The WMDP Benchmark: Measuring and Reducing Malicious Use with Unlearning