acceptodds
Under review as a conference paper at ICLR 2027

Crossing Lethe: Steering Probability Flows for Data Unlearning

Abstract

Data unlearning (the targeted removal of specific training examples from a deployed model) is increasingly important for privacy compliance, user rights, and safe model updating. While unlearning has been actively studied for energy-based models and diffusion models, comparatively little work addresses flow matching models, in part because it is unclear how removing data should alter the learned dynamics. Flow matching trains continuous generative models by learning a time-dependent velocity field whose estimates aggregate contributions from all training samples, making post-training selective removal nontrivial. We propose a flow decomposition framework that adapts to the level of data access and operational constraints. We first introduce Loss-level Flow Decomposition (LFD), that explicitly subtracting the forget-set contribution. Importantly, LFD serves as a global optimizer that obviates the need to retrain the model. LFD provides high fidelity in standard settings where both retain and forget datasets are available. We then investigate privacy-restricted scenarios where explicit access to forget samples is prohibited or governed by strict constraints. Specifically, we introduce an Energy-Based Reformulation (EBR) that leverages a proxy classifier rather than direct forget data samples. We rigorously analyze both these approaches. We also validate our framework on several datasets, demonstrating that these methods effectively balance unlearning performance with the preservation of model fidelity across both standard and restricted data regimes.

Primary area:
generative models
Keywords:
Data Unlearning, Flow Matching, Generative Models, Deep Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ebdROpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies data unlearning in flow matching generative models and proposes two methods, LFD and EBR. LFD uses loss decomposition and importance sampling to estimate the retain-only training objective when both retain and forget data are available. EBR instead uses proxy classifier scores to reweight the training distribution when direct access to forget samples is unavailable. The paper provides theoretical analyses of both methods and evaluates them on CelebA identity removal and MNIST digit-class unlearning.

Critical strengths
  1. The paper addresses a relevant problem by studying data unlearning in flow matching models and considering different levels of access to forget data.
  2. The paper provides mathematical derivations and includes exact retraining as an experimental reference. The results on CelebA and MNIST show that the proposed methods can reduce the generation of targeted content while preserving some retained generative capability.
Critical weaknesses
  1. The methodological novelty is limited, with substantial overlap with existing work. The loss decomposition and importance sampling mechanism in LFD closely resemble SISS (ICLR 2025), and the paper itself acknowledges their mathematical connection in Appendix E. The energy-reweighted objective of EBR directly corresponds to existing Energy-Weighted Flow Matching formulations (ICLR 2025) and is also closely related to the proxy-based unlearning approach in ContinualFlow (2025). Although these techniques are adapted to the proposed setting, the paper does not sufficiently establish its distinct technical contributions, particularly compared with EFM and related methods.
  2. The experimental evaluation is insufficient to support the main claims. LFD and EBR are evaluated separately on CelebA and MNIST, without cross-dataset validation or experiments on more challenging generative models. The CelebA experiments lack qualitative generation comparisons, and all reported results use Weighted LFD with an additional forget-ascent term rather than the theoretically unbiased Exact LFD. This makes it difficult to isolate the contribution of the proposed loss decomposition. On MNIST, the results remain substantially worse than exact retraining. For example, Table 2 reports a forgotten-class fraction of 10.7% for EBR compared with 1.0% for retraining. Furthermore, the ContinualFlow comparison relies on a local reimplementation rather than the original image-domain system. The presentation is also unnecessarily complicated, making the methodology and actual contributions difficult to follow.

Alberti et al., Data Unlearning in Diffusion Models (SISS), ICLR 2025 Zhang et al., Energy-Weighted Flow Matching for Offline Reinforcement Learning, ICLR 2025 Simone et al., ContinualFlow: Learning and Unlearning with Neural Flow Matching, ICML 2025

Questions for authors
  1. Could the authors clarify the main technical differences between LFD and SISS, and between EBR and existing Energy-Weighted Flow Matching or ContinualFlow methods? In particular, what new theoretical or algorithmic contributions do EBR's Eq. (4) and Eq. (6) provide beyond existing energy-weighted objectives?
  2. Why are results for Exact LFD with (\rho=0) not reported on CelebA? Since all main experiments use an additional forget-ascent term, how can the improvements be attributed to the proposed unbiased loss decomposition rather than this extra update? Could the authors provide the corresponding ablation and qualitative face-generation results?
  3. Could the authors explain the substantial performance gap between EBR and exact retraining, and provide sensitivity analyses for proxy quality or weighting parameters? Also, why is the comparison limited to a local reimplementation of ContinualFlow instead of a more faithful comparison with its original image-domain method?
3 comments
AuthorsOpenReview ↗

Rebuttal Part-1

We thank the reviewer for their meaningful feedback, specifically for highlighting our two-regime formulation and exact-retrain benchmarks, and for pointing us to EFM (Zhang et al., ICLR 2025).

Weakness 1: Novelty vs. SISS, EFM & ContinualFlow

We will cite EFM, present general Boltzmann-weighted flow matching (Eqs. 4 and 6) as background, and remove "global optimizer" from the Abstract so our genuine contributions stand out clearly. In particular, unlike ContinualFlow—which freezes the unedited retain+forget base generator (keeping forgotten data intact in its weights) and trains a second flow network on top (doubling test-time ODE sampling cost)—EBR overwrites the single base generator in place (1× sampling cost) using exact prior-corrected density ratios.

Q1. (LFD vs. SISS) SISS formulates importance-sampled unlearning for score-based diffusion SDEs; under linear Gaussian bridges xt=(1−t)x0+tzx_t = (1 - t)x_0 + t z (Appendix E), LFD shares SISS's defensive-mixture weights while regressing deterministic ODE velocities utθu_t^\theta directly—inducing a ((1−t)/t)2((1 - t)/t)^2 time weight relative to unweighted score matching without a score network or stochastic SDE sampler—and adds our O(πF)O(\pi_F) gradient-dilution diagnosis and ρ∝Nf\rho \propto N_f size-scaling rule. (EBR vs. EFM & ContinualFlow) EFM assumes a known offline-RL reward R(z)R(z), whereas unlearning provides only an imperfect classifier r^(z)\hat r(z) whose training prior differs from the deployment prior πR\pi_R; Proposition B.3 proves that prior-shift odds correction at β=1\beta = 1 recovers the exact retain law q0(⋅∣R=1)q_0(\cdot \mid R = 1) and splits squared density-ratio error into irreducible feature insufficiency (E[Var(r(Z)∣S)]/πR2\mathbb E[\mathrm{Var}(r(Z) \mid S)]/\pi_R^2) plus calibration error, while updating one noise-to-data generator in place rather than ContinualFlow's heuristic σ(−γ⋅logit)\sigma(-\gamma \cdot \text{logit}) two-model q0→q1q_0 \to q_1 wrapper.

Prior WorkKey Limitation for Flow UnlearningOur Distinct ContributionPractical Payoff
SISS (ICLR 2025)Score-based diffusion SDEsVelocity ODE LFD + O(πF)O(\pi_F) diagnosis + ρ∝Nf\rho \propto N_fNative ODEs; zero-shot size scaling
EFM (ICLR 2025)Assumes known RL reward R(z)R(z)Prop. B.3: β=1\beta = 1 retain recovery & L2L^2 error splitExact prior fix; insufficiency vs. calibration
ContinualFlow (2025)2-model wrapper (base kept; 2× cost)In-place 1-model sanitization via Bayes ratios1× sampling cost; base weights overwritten

Each regime contributes distinct unlearning theory, scaling rules, and single-model deployment beyond SISS, EFM, and ContinualFlow.

Weakness 2: Exact LFD, Cross-Dataset Scope, Faces & MNIST Gap

We will streamline the presentation, separate Exact LFD (ρ=0\rho = 0) from Eq. (2)'s rescaled forget step in Section 5, and add five-seed CIFAR-10 class-removal experiments where both LFD and EBR succeed on the same base generator—alongside the exact sample-starvation boundary showing why rare identity deletion requires LFD.

Q2. (Exact LFD & attribution) On CelebA (Gap = Method − Gold on target-identity FaceNet reconstruction cosine across 5 seeds, 256 states/step), the 180 forget images represent πF≈10−3\pi_F \approx 10^{-3} of training data, making LFD's raw forget gradient gFg_F ~400× smaller in norm than its retain gradient gRg_R despite aligning at cosine +0.981+0.981 with direct forget ascent. Consequently, Exact LFD (ρ=0\rho = 0) stalls like Naive retain-only fine-tuning, whereas Eq. (2) (rescaling gFg_F to norm ρ∥gR∥2\rho \|g_R\|_2) drives >1,000× faster convergence to Gold: it matches a direct 128-retain + 128-forget split (Retain + Forget Ascent; paired 95% CIs span 0 across steps 12–60, seed-17 step-12 FID 30.1 vs. 31.5), centers single-identity runs (Nf=30N_f = 30) on Gold (−0.0007 vs. −0.0248 unscaled) via ρ(Nf)=ρ0(Nf/Nf,0)\rho(N_f) = \rho_0(N_f / N_{f,0}), and on 5-seed CIFAR-10 airplane removal cuts airplanes from 9.11% (Base) and 5.67% (Naive) to 4.18% (p=0.031p = 0.031, retain FID 11.55 vs. 11.50). (Qualitative faces) Because OpenReview comments are text-only, we will add our paired 7-column CelebA reconstruction figure (Original, Base, Gold, Naive, Exact LFD, Retain + Forget Ascent, Weighted LFD from identical noise seeds) to the revised paper; visually, Weighted LFD erases target-identity traits to match Gold while preserving pose, lighting, and non-deleted identities (step-12 retained cosine 0.2718 vs. 0.2730 for Gold).

AuthorsOpenReview ↗

Rebuttal Part-2

CelebA Joint-Six (Nf=180N_f = 180, 256 states/step, 5 seeds)Target Cosine @ Step 12Gap to Gold @ Step 12Gap to Gold @ Step 60Retained FID @ Step 60 ↓
Base model (before unlearning)0.2842+0.0051+0.005127.8 (step 0)
Exact retrain from scratch (Gold, 30k steps)0.27910.00000.000027.7
Retain-only fine-tuning (Naive)0.2855+0.0064 ± 0.0056+0.0056 ± 0.005325.4 (step 132)
Exact LFD-MC (ρ=0\rho = 0)0.2861+0.0070 ± 0.0033+0.0035 ± 0.003126.8
Weighted LFD-MC (ρ=0.05\rho = 0.05)0.2815+0.0024 ± 0.0031−0.0073 ± 0.002226.6
Weighted LFD-MC (ρ=0.10\rho = 0.10)0.2764−0.0027 ± 0.0028−0.0195 ± 0.004426.4
Retain + Forget Ascent (128 retain + 128 forget, ρ=0.10\rho = 0.10)0.2738−0.0053 ± 0.0071−0.0175 ± 0.002228.9

Eq. (2) supplies the active repulsion needed for rapid erasure; Retain + Forget Ascent uses pre-split sets, while LFD matches it on unsplit (Pdata,Pforget)(P_{\text{data}}, P_{\text{forget}}).

Q3. (Why EBR stops short of 1.0% Gold) Without real retain images, classifier scores can only reweight the base generator's imperfect 65,536-sample synthetic bank q0q_0: even a 0%-error Oracle filter (using the test evaluator itself to strip 100% of odd digits from the bank) leaves 9.2% generated odd digits at 300 steps, and 1,200-step EBR plateaus at 6.1%—demonstrating an inherent ceiling shared by all data-free methods that rely solely on classifier scores. (Proxy sensitivity & cross-dataset scope) Dividing proxy logits by 4 (making probabilities uncertain at 99.0% accuracy) degrades β=1\beta = 1 to 16.6%, whereas sharpening to β=4\beta = 4 cuts leakage to 6.7% (beating hard filtering's 9.9%); downsampling 28×2828 \times 28 digits 4×4\times to blurry 7×77 \times 7 inputs (97.5% proxy accuracy, testing Proposition B.3's lossy view S=g(Z)S = g(Z)) raises leakage only slightly to 11.2% (10.2% Hard). On 5-seed CIFAR-10, data-free EBR cuts airplanes from 9.41% (uniform bank, FID 16.94) to 5.67% (FID 15.87, p=0.031p = 0.031); conversely, on single-ID CelebA (πF≈1.8×10−4\pi_F \approx 1.8 \times 10^{-4}), a 65,536-sample bank contains only 4 target-person images, starving any bank-reweighting method and confirming rare-instance removal requires LFD. (ContinualFlow) Simone et al. (2025) achieve 0.05% odd leakage using an unreleased convolutional autoencoder + latent flow (releasing only 2D toy code). Matching Appendix G, we will rewrite Section 7.2 and Table 2's caption as "Pixel-space ContinualFlow objective", clarifying that direct q0→q1q_0 \to q_1 image-to-image interpolation stalls in pixel space without an autoencoder whereas EBR succeeds in pixel space inside one model.

MNIST Odd-Digit Removal (65k synthetic bank)SeedsStepsOdd Digits Left ↓Retain Feature FID ↓
Base model (generates ~50% odd, ~50% even)1049.3%357.5
Exact retrain on real even digits (Gold)1117.5k1.0%1.8
Uniform bank fine-tuning (β=0\beta = 0)530048.5 ± 3.7%377.5 ± 42.7
EBR (β=1\beta = 1, 99.0% proxy)530010.7 ± 1.0%51.7 ± 15.3
Hard filter (99.0% proxy)230010.3%61.2
Oracle classifier filter (0% error on bank)23009.2%63.9
EBR (β=1\beta = 1, 99.0% proxy, longer budget)16006.6%27.8
Under-confident proxy (logits ÷4\div 4), EBR (β=1\beta = 1)130016.6%83.5
Under-confident proxy (logits ÷4\div 4), Hard filter13009.9%60.2
Under-confident proxy (logits ÷4\div 4), EBR (β=4\beta = 4)13006.7%73.5
7×77 \times 7-downsampled proxy (97.5% acc.), EBR (β=1\beta = 1)130011.2%65.6
7×77 \times 7-downsampled proxy (97.5% acc.), Hard filter130010.2%67.6

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts to address my concerns. However, three important questions remain:

1. Methodological novelty. The authors acknowledge that LFD shares its loss decomposition and importance-sampling mechanism with SISS, while EBR's energy-weighted objective follows existing EFM formulations. Although the velocity parameterization, prior correction, and single-model deployment provide useful distinctions, could the authors clarify what genuinely new algorithmic or theoretical contributions these methods introduce beyond adapting established techniques to the unlearning setting? In particular, how does Proposition B.3 go beyond standard Bayesian identities and conditional-expectation decomposition?

2. LFD's actual contribution and experimental attribution. The new ablations show that Exact LFD provides limited improvement over naive retain-only fine-tuning, while Weighted LFD achieves its gains primarily through the additional norm-controlled forget-ascent term, which resembles existing strategies in SISS. Moreover, Retain + Forget Ascent achieves comparable results. What evidence demonstrates that LFD's proposed decomposition, rather than the additional forget-ascent mechanism, is responsible for the reported improvements? Does Weighted LFD offer a distinct computational or optimization advantage over this simpler baseline?

3. EBR's theoretical claims and experimental evidence. Proposition B.3 establishes recovery of the retain-conditional distribution of the synthetic generator, which need not match the real-data retraining distribution. Furthermore, the oracle-filter experiment's 9.2% residual leakage after 300 updates does not establish an inherent limitation of classifier-only methods, as optimization and approximation errors remain possible. Could the authors clarify these theoretical distinctions and explain how the remaining gap to exact retraining should be interpreted? Given the acknowledged limitations of the pixel-space ContinualFlow reimplementation, what conclusions about EBR's comparative effectiveness remain supported?

I appreciate the additional evidence and would welcome clarification of these points before finalizing my assessment.

Reviewer_FbVoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies empirical unlearning for flow-matching generators under two data-access regimes. Loss-level Flow Decomposition (LFD) uses a full-minus-forget risk identity and a defensive-mixture estimator to estimate the retain-only flow-matching objective, and a Rao-Blackwellized version removes proposal-branch randomness. The evaluated CelebA implementation adds a norm-controlled forget-ascent update. Energy-Based Reweighting (EBR) instead fine-tunes on synthetic endpoints weighted by a frozen retain-membership proxy. The paper provides population-level identities and evaluates identity deletion on CelebA and odd/even digit removal on MNIST, using exact retraining as a behavioral reference.

Critical strengths
  1. The explicit separation of retain/forget access from proxy-only access is useful. The estimator derivations and the discussion of calibration, proxy insufficiency, and finite-batch bias make the assumptions more inspectable than a purely heuristic unlearning recipe. 
  2. The experiments include exact-retraining references, reconstruction-based identity measurements, retained-quality measurements, and five unlearning seeds.
Critical weaknesses
  1. The central theoretical result does not establish the behavior of the method responsible for the empirical gains. The paper does not report a matched CelebA comparison of exact LFD, direct retain-only fine-tuning, and direct retain-only fine-tuning plus the same norm-controlled forget-ascent term. Appendix D.2 also reports that the single-identity runs overshoot their matched-gold reference at the first measured checkpoint. A lower target-similarity score is therefore not sufficient evidence of the stated retraining-equivalence objective. The limitations acknowledge this issue, but a deletion-size-aware operating rule and an empirical demonstration of the estimator's distinct contribution are still missing. 
  2. The experimental evidence is too limited to substantiate the claimed quality-preserving approximation to retraining across the two regimes. EBR's oracle identity recovers the retain conditional of the accessible law q0. When q0 is a generated bank, this is not automatically the empirical retain-data law used for retraining. The missing link between these targets and the limited robustness evidence leave the practical unlearning claim insufficiently established.
Questions for authors
  1. Can you provide a matched-budget CelebA comparison of exact LFD (rho = 0), direct retain-only fine-tuning, and retain-only fine-tuning with the same norm-controlled forget-ascent term? 
  2. How are rho and stopping time selected for a new deletion request without using its exact-retraining model, and can the same selection rule avoid the overshoot reported for single-identity deletion? 
  3. With the same synthetic bank, proxy, and update budget, how does EBR compare with hard-filtered bank distillation? Can you report retained-quality uncertainty across seeds and explain how generator-bank mismatch contributes to the remaining gap to gold?
2 comments
AuthorsOpenReview ↗

Rebuttal Part-1

We thank the reviewer for their meaningful feedback, specifically for highlighting our explicit vs. proxy-only separation and 5-seed exact-retraining protocol.

Weakness 1: Theory–Practice Link & Overshoot Rule

We agree that signed proximity to exact retraining - rather than driving target similarity arbitrarily low - is the right unlearning objective. Exact LFD (ρ=0\rho = 0) provides an unbiased retain-risk estimator via a defensive bridge mixture (Theorem 5.3), decomposing each update into a retain gradient gRg_R and a forget gradient gFg_F, but stalls over short budgets because gFg_F is diluted by the small forget prior πF\pi_F. Fast short-budget convergence comes from Eq. (2), which rescales gFg_F to norm ρ∥gR∥2\rho \|g_R\|_2; we will clearly separate Exact LFD (ρ=0\rho = 0) from Eq. (2) in the Abstract and Section 5, and show below how scaling ρ\rho with forget-set size prevents single-identity overshoot.

Q1. (Exact LFD vs. Naive) Across five CelebA seeds at a matched 256 bridge states (256 forward passes) per step - measuring target-identity FaceNet reconstruction cosine and Gap = Method - Gold with 95% Student-tt CIs (copy-detection similarity via SSCD ranks all methods identically)—Exact LFD-MC (ρ=0\rho = 0) tracks Naive retain-only fine-tuning at 1.9×1.9\times higher minibatch retain-gradient variance (7.9 vs. 4.2). (Retain + Forget Ascent) Applying Eq. (2) directly to 128 pre-separated retain states and 128 forget states (Retain + Forget Ascent) produces a trajectory statistically indistinguishable from Weighted LFD-MC across steps 12–60 (all paired 95% CIs span 0). At step 12, Weighted LFD-MC keeps non-deleted retained-identity cosine within −0.0012 ± 0.0061 of Gold, with seed-17 10k-sample Retained FID of 30.1 vs. 31.5 for Retain + Forget Ascent (and 26.4 vs. 28.9 at step 60).

CelebA Joint-Six (Nf=180N_f = 180, 256 states/step, 5 seeds)Target Cosine @ Step 12Gap to Gold @ Step 12Gap to Gold @ Step 60Retained FID @ Step 60 ↓
Base model (before unlearning)0.2842+0.0051+0.005127.8 (step 0)
Exact retrain from scratch (Gold, 30k steps)0.27910.00000.000027.7
Retain-only fine-tuning (Naive)0.2855+0.0064 ± 0.0056+0.0056 ± 0.005325.4 (step 132)
Exact LFD-MC (ρ=0\rho = 0)0.2861+0.0070 ± 0.0033+0.0035 ± 0.003126.8
Weighted LFD-MC (ρ=0.05\rho = 0.05)0.2815+0.0024 ± 0.0031−0.0073 ± 0.002226.6
Weighted LFD-MC (ρ=0.10\rho = 0.10)0.2764−0.0027 ± 0.0028−0.0195 ± 0.004426.4
Retain + Forget Ascent (128 retain + 128 forget, ρ=0.10\rho = 0.10)0.2738−0.0053 ± 0.0071−0.0175 ± 0.002228.9

This cleanly separates both roles: Eq. (2) drives rapid short-budget erasure, Retain + Forget Ascent works equally well when retain and forget indices are pre-split, and LFD matches it directly from an unsplit PdataP_{\text{data}} stream plus PforgetP_{\text{forget}}.

Q2. (Why fixed ρ\rho overshoots) Because Eq. (2) normalizes gF/norm(gF)g_F/norm(g_F) to unit norm, the forget update has total norm ρ⋅norm(gR)\rho \cdot norm(g_R) regardless of how many images NfN_f are being forgotten—concentrating 180/30=6×180 / 30 = 6\times stronger repulsion per image onto a single-identity request (Nf=30N_f = 30) than our joint six-identity request (Nf,0=180N_{f,0} = 180). (Selecting ρ\rho & stopping) Calibrating (ρ0,T0)=(0.10,12)(\rho_0, T_0) = (0.10, 12) once on a reference deletion (exploratory on Nf,0=180N_{f,0} = 180 here; one held-out validation deletion in practice) and scaling ρ(Nf)=ρ0(Nf/Nf,0)\rho(N_f) = \rho_0(N_f / N_{f,0}) keeps repulsion per image constant (ρ=0.0167\rho = 0.0167 for Nf=30N_f = 30) at the same fixed step count T0=12T_0 = 12, evaluated below against scratch-retrained single-identity Gold models across 3 seeds.

Single-Identity CelebA (Nf=30N_f = 30, 6 IDs, 3 gold seeds)Gap to Gold @ Step 12Gap to Gold @ Step 24IDs Below Gold @ Step 12
Unscaled Weighted LFD-MC (ρ=0.10\rho = 0.10)−0.0248 ± 0.0057−0.0394 ± 0.01426 of 6 (overshoots)
Retain-only fine-tuning (Naive)+0.0057 ± 0.0133+0.0023 ± 0.00060 of 6 (under-forgets)
Size-scaled Weighted LFD-MC (ρ=0.0167\rho = 0.0167)−0.0007 ± 0.0059−0.0051 ± 0.00303 of 6 (centered)

Scaling ρ∝Nf\rho \propto N_f centers single-identity runs on Gold at T0=12T_0 = 12 without training per-request Gold models, while mild drift by step 24 (−0.0051) confirms that pairing ρ(Nf)\rho(N_f) with fixed-T0T_0 stopping is essential.

AuthorsOpenReview ↗

Rebuttal Part-2

Weakness 2: Generator-Bank Mismatch & Quality Uncertainty

We agree that because EBR has no access to real training images, its oracle target q0(⋅∣R=1)q_0(\cdot \mid R = 1) is the retain subset of the base generator's synthetic sample bank q0q_0, which differs from the real retain dataset PretainP_{\text{retain}} (used to train Gold) whenever the base generator is imperfect. Testing hard filtering, a 0%-error Oracle filter, degraded proxies, and 600–1,200 step budgets on the same 65,536-image bank isolates each error source.

Q3. (5-seed quality CIs & CIFAR-10) Across five MNIST seeds, EBR preserves class balance among the five kept digits (Retain TV: total variation distance to uniform kept digits) and lowers evaluator feature-space FID to real kept digits on both parities (forgetting even leaves 8.8 ± 1.4% vs. 0.8% Gold, Retain FID 43.1 ± 19.0, Retain TV 0.078 ± 0.036). On 5-seed CIFAR-10 airplane removal (scoring a synthetic bank with a ResNet-20 proxy), EBR likewise cuts airplanes from 9.41% (uniform bank distillation, FID 16.94) to 5.67% (FID 15.87, p=0.031p = 0.031). (Hard filtering & proxy ablations) On our well-calibrated 99.0% MNIST proxy, near-binary scores make β=1\beta = 1 match hard filtering (r^≥0.5\hat r \ge 0.5); when proxy logits are divided by 4 (making scores uncertain while keeping 99.0% accuracy), setting β=4\beta = 4 to match expected retain mass on q0q_0 beats hard filtering on leakage (concentrating weight onto 45.9% vs. 51.4% effective bank samples, which raises FID/TV); and downsampling 28×2828 \times 28 digits 4×4\times via average pooling to blurry 7×77 \times 7 images (dropping proxy accuracy to 97.5% to test Proposition B.3's lossy view S=g(Z)S = g(Z)) degrades EBR only slightly from 10.7% to 11.2%.

MNIST Odd-Digit Removal (65k synthetic bank)SeedsStepsOdd Digits Left ↓Retain Feature FID ↓Retain Class TV ↓
Base model (~50% odd, ~50% even)1049.3%357.50.014
Exact retrain on real even digits (Gold)1117.5k1.0%1.80.013
Uniform bank fine-tuning (β=0\beta = 0)530048.5 ± 3.7%377.5 ± 42.70.072 ± 0.024
EBR (β=1\beta = 1, 99.0% proxy)530010.7 ± 1.0%51.7 ± 15.30.080 ± 0.041
Hard filter (99.0% proxy)230010.3%61.20.108
Oracle filter (evaluator labels, 0% bank odd)23009.2%63.90.122
EBR (β=1\beta = 1, 99.0% proxy, 600 steps)16006.6%27.80.072
EBR (β=1\beta = 1, 99.0% proxy, 1200 steps)112006.1%36.40.081
Under-confident (logits ÷4\div 4), EBR (β=1\beta = 1)130016.6%83.50.069
Under-confident (logits ÷4\div 4), Hard filter13009.9%60.20.092
Under-confident (logits ÷4\div 4), EBR (β=4\beta = 4)13006.7%73.50.127
7×77 \times 7-downsampled proxy (97.5% acc.), EBR (β=1\beta = 1)130011.2%65.60.083
7×77 \times 7-downsampled proxy (97.5% acc.), Hard filter130010.2%67.60.092

Even when the test evaluator itself removes 100% of odd digits from the bank (Oracle filter), 300-step fine-tuning still generates 9.2% odd digits and 1,200-step EBR plateaus at 6.1%: proxy error explains ~1.5 points and step budget ~4.6 points, leaving a ~5-point theoretical floor shared by all data-free classifier methods that reweight synthetic q0q_0 rather than real PretainP_{\text{retain}}.

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

This work proposes a machine unlearning framework for flow matching-based generative models that removes the influence of specific training data without requiring full retraining. It introduces two methods: Loss-level Flow Decomposition (LFD), which subtracts the forget-set contribution from the training objective when retain and forget data are accessible, and Energy-Based Reweighting (EBR), which uses a proxy classifier to reweight the training distribution when direct access to forget data is restricted. Experiments demonstrate that LFD can remove specific identities from CelebA while preserving generation quality, whereas EBR reduces the generation of targeted digit classes in MNIST.

Critical strengths

S1. Meaningful theoretical extension to flow matching.

The paper extends existing unlearning principles from diffusion models to velocity-field learning in flow matching, providing a theoretical formulation that clarifies the relationship between these two generative modeling frameworks.

S2. Insightful analysis of information constraints in unlearning.

The paper provides a theoretically grounded analysis of how access to forget data and the quality of proxy information affect the attainable target distribution. In particular, the oracle-equivalence and proxy-insufficiency analyses offer useful insights into the limitations of unlearning under restricted data access.

Critical weaknesses

Weaknesses

W1. Unclear algorithmic advantage and disconnect between theory and experiments.

Although LFD provides an unbiased estimator of the retain-only flow-matching objective, direct retain-only fine-tuning also yields an unbiased estimator with a single model evaluation per sampled state. Moreover, the theoretical guarantees apply to Exact LFD (ρ=0\rho=0), whereas all reported CelebA experiments use Weighted LFD (ρ=0.1\rho=0.1), which introduces an additional forget-gradient ascent component. Without a compute-matched comparison, it remains unclear whether the observed improvements arise from the proposed decomposition or the additional gradient update, and what practical advantage LFD offers over simpler alternatives.

W2. Insufficient experimental validation and limited generalizability.

The experiments are restricted to relatively small flow-matching models and limited unlearning settings, with LFD evaluated only on CelebA identity removal and EBR on MNIST parity removal. In particular, EBR lacks comparisons with simpler proxy-based alternatives, such as hard filtering, despite using a highly accurate proxy classifier (99%). The absence of evaluations on more challenging datasets, larger flow-matching models, and diverse forgetting scenarios makes it difficult to assess the framework's generalizability and practical significance.

Questions for authors

Q1. Could the authors provide a compute-matched comparison among direct retain-only fine-tuning, Exact LFD ((\rho=0)), and Weighted LFD ((\rho=0.1))? In particular, comparing gradient variance, convergence behavior, and forgetting–retention trade-offs would help clarify the practical benefit of the proposed estimator.

Q2. How much of Weighted LFD's empirical improvement can be attributed to the theoretically justified estimator versus the additional forget-gradient ascent term? An ablation varying (\rho), including (\rho=0), would help establish this distinction.

Q3. How does EBR compare with simpler proxy-based alternatives, such as hard filtering and uniform fine-tuning on the generated endpoint bank? It would also be valuable to examine sensitivity to proxy accuracy and calibration, particularly when retain/forget membership is uncertain.

Q4. Could the authors expand the evaluation to a more challenging flow-matching setting or provide additional generalization evidence, such as varying forget-set sizes, sequential unlearning requests, or a larger text-conditioned model? In addition, could the authors clarify what theoretical or empirical advantages LFD provides over SISS specifically in the deterministic flow-matching setting?

2 comments
AuthorsOpenReview ↗

Rebuttal Part-1

We thank the reviewer for their meaningful feedback, specifically for highlighting our velocity-field formulation and proxy-insufficiency analysis.

Weakness 1: Algorithmic Advantage & Attribution

We agree that our fast short-budget unlearning comes from the norm-calibrated forget-ascent term in Eq. (2) rather than the unbiased Exact LFD estimator (ρ=0\rho = 0), and we will make this distinction explicit in the Abstract and Section 5. Because Conditional Flow Matching (CFM) is linear in the data distribution PdataP_{\text{data}} (Theorem 5.1), our defensive bridge mixture (Theorem 5.3) matches retain-only fine-tuning in expectation while decomposing each update into an unbiased retain gradient gRg_R and an explicit forget gradient gFg_F. Eq. (2) then rescales gFg_F so its norm is a fixed fraction ρ\rho of ∥gR∥2\|g_R\|_2, overcoming rare-set gradient dilution.

Q1. (Variance) All CelebA methods below are compute-matched at 256 bridge states (256 forward passes) per step across five seeds. Across 64 shared minibatches, Exact LFD's gradient aligns closely with Naive retain-only fine-tuning (cosine 0.91–0.96) at 1.9×1.9\times relative variance (variance over squared mean norm: 7.9 vs. 4.2), without weight blow-up. (Convergence & retention) Measuring target-identity FaceNet reconstruction cosine (Gap = Method − Gold, where 0 matches scratch retraining and < 0 overshoots it, with 95% tt-CIs), Weighted LFD reaches Gold in 12 steps (>1,000× fewer steps than retraining), keeping retained-identity cosine within 0.002 of Gold, 0% false matches, and seed-17 10k-sample FID near Gold (30.1 Weighted LFD and 29.2 Exact LFD at step 12; step-60 FID below).

CelebA Method (256 states/step, 5 seeds)Target Cosine @ Step 12Gap to Gold @ Step 12Retained FID @ Step 60 ↓
Base model (before unlearning)0.2842+0.005127.8 (step 0)
Exact retrain (Gold, 30k steps)0.27910.000027.7
Retain-only fine-tuning (Naive)0.2855+0.0064 ± 0.005625.4 (step 132)
Exact LFD (ρ=0\rho = 0)0.2861+0.0070 ± 0.003326.8
Weighted LFD (ρ=0.10\rho = 0.10)0.2764−0.0027 ± 0.002826.4

At matched per-step compute, Weighted LFD reaches exact retraining in 12 steps with stable updates and preserved retained quality.

Q2. (Attribution & ρ\rho sweep) Because the 6 deleted identities are only πF≈10−3\pi_F \approx 10^{-3} of CelebA (180 of 171,306 images), gFg_F points along direct forget ascent (cosine +0.981+0.981) yet is ~400× smaller in norm than gRg_R (4.3×10−44.3 \times 10^{-4} vs. 0.170.17), so passive ρ=0\rho = 0 stalls while ρ>0\rho > 0 smoothly controls erasure. (Direct split comparison) Spending the same 256 passes/step on 128 pre-separated retain images (gRg_R) and 128 forget images (gFg_F) under Eq. (2) (Retain + Forget Ascent, step-12 FID 31.5) matches Weighted LFD across steps 12–60 (all paired 95% CIs span 0).

CelebA ρ\rho-Sweep & Direct SplitGap to Gold @ Step 12Gap to Gold @ Step 60Retained FID @ Step 60 ↓
Exact LFD (ρ=0\rho = 0)+0.0070 ± 0.0033+0.0035 ± 0.003126.8
Weighted LFD (ρ=0.05\rho = 0.05)+0.0024 ± 0.0031−0.0073 ± 0.002226.6
Weighted LFD (ρ=0.10\rho = 0.10)−0.0027 ± 0.0028−0.0195 ± 0.004426.4
Weighted LFD (ρ=0.20\rho = 0.20)−0.0139 ± 0.0036−0.0615 ± 0.021227.2
Retain + Forget Ascent (128 + 128, ρ=0.10\rho = 0.10)−0.0053 ± 0.0071−0.0175 ± 0.002228.9

Eq. (2) drives fast short-budget erasure: Retain + Forget Ascent works equally well on pre-split sets, while LFD matches it when sampling from unsplit PdataP_{\text{data}} plus PforgetP_{\text{forget}}.

AuthorsOpenReview ↗

Rebuttal Part-2

Weakness 2: Baselines & Generalizability

We expand validation with 5-seed CIFAR-10 class unlearning for both LFD and EBR, MNIST bank and oracle baselines, degraded-proxy accuracy and calibration tests, and single-identity size scaling.

Q3. (Bank baselines) Without real data, EBR fine-tunes on a 65,536-image synthetic bank q0q_0 sampled from the base model: uniform fine-tuning (β=0\beta = 0) gives no forgetting, whereas our confident 99.0%-accurate parity classifier outputs near-binary scores so EBR (β=1\beta = 1) matches hard thresholding (r^≥0.5\hat r \ge 0.5). (Proxy calibration & accuracy) Dividing proxy logits by 4 keeps 99.0% accuracy while making probabilities uncertain: β=1\beta = 1 leaves 16.6% odd digits, hard filtering gets 9.9%, and sharpening EBR to β=4\beta = 4 (r^(z)4\hat r(z)^4) cuts leakage to 6.7% (concentrating weight onto fewer bank images raises feature-space FID to real even digits). To test lower proxy accuracy (Proposition B.3's lossy view S=g(Z)S = g(Z)), we downsampled 28×2828 \times 28 digits 4×4\times via average pooling to blurry 7×77 \times 7 images, dropping proxy accuracy from 99.0% to 97.5%; EBR degrades gracefully from 10.7% to 11.2%. (Classifier ceiling) Even a 0%-error Oracle filter (using the test evaluator itself to remove 100% of odd digits from the bank) leaves 9.2% odd outputs at 300 steps, showing an inherent synthetic-bank ceiling shared by all classifier-score methods without real retain data.

MNIST Odd Removal (65k synthetic bank)SeedsStepsOdd Digits Left ↓Retain Feature FID ↓
Base model (~50% odd, ~50% even)1049.3%357.5
Exact retrain on real even digits (Gold)1117.5k1.0%1.8
Uniform bank fine-tuning (β=0\beta = 0)530048.5 ± 3.7%377.5 ± 42.7
EBR (β=1\beta = 1, 99.0% proxy)530010.7 ± 1.0%51.7 ± 15.3
Hard filter (99.0% proxy)230010.3%61.2
Oracle filter (0% error on bank)23009.2%63.9
EBR (β=1\beta = 1, 99.0% proxy, 600 steps)16006.6%27.8
Under-confident (logits ÷4\div 4), EBR (β=1\beta = 1)130016.6%83.5
Under-confident (logits ÷4\div 4), Hard filter13009.9%60.2
Under-confident (logits ÷4\div 4), EBR (β=4\beta = 4)13006.7%73.5
7×77 \times 7-downsampled (97.5% acc.), EBR (β=1\beta = 1)130011.2%65.6
7×77 \times 7-downsampled (97.5% acc.), Hard filter130010.2%67.6

EBR matches hard filtering when proxies are sharp, degrades gracefully on lower-accuracy 7×77 \times 7 proxies, beats hard filtering under miscalibration, and approaches the synthetic-bank limit.

Q4. (Generalization to CIFAR-10) Removing airplanes (5,000 images) from a pretrained CIFAR-10 flow model over 300 steps across 5 seeds (p=0.031p = 0.031), Weighted LFD cuts airplanes from 9.11% (Base) and 5.67% (Naive retain-only) to 4.18% at flat retain FID (11.55 vs. 11.50), while data-free EBR cuts synthetic-bank distillation from 9.41% to 5.67% (FID 16.94 to 15.87). (Size scaling) Because Eq. (2) normalizes gFg_F to unit norm, scaling ρ(Nf)=ρ0(Nf/Nf,0)\rho(N_f) = \rho_0(N_f / N_{f,0}) keeps repulsion per forgotten image constant (0.10×30/180=0.01670.10 \times 30/180 = 0.0167 from 6 identities to 1). (LFD vs. SISS) Under Gaussian bridges xt=(1−t)x0+tzx_t = (1 - t)x_0 + t z, LFD shares SISS's mixture weights while regressing ODE velocities directly—a ((1−t)/t)2((1 - t)/t)^2 time weight vs. unweighted score matching without a score network or SDE sampler—and adds ρ∝Nf\rho \propto N_f scaling; sequential and text flows are future work.

Single-Identity CelebA (Nf=30N_f = 30, 3 gold seeds)Gap to Gold @ Step 12Gap to Gold @ Step 24IDs Below Gold @ Step 12
Unscaled Weighted LFD (ρ=0.10\rho = 0.10)−0.0248 ± 0.0057−0.0394 ± 0.01426 of 6 (overshoots)
Retain-only fine-tuning (Naive)+0.0057 ± 0.0133+0.0023 ± 0.00060 of 6 (under-forgets)
Size-scaled Weighted LFD (ρ=0.0167\rho = 0.0167)−0.0007 ± 0.0059−0.0051 ± 0.00303 of 6 (centered)

Scaling ρ∝Nf\rho \propto N_f eliminates single-identity overshoot at step 12 without per-request retuning.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.