{"id":"ed777686-9f18-41e3-b7df-80e98c1e6f3e","arxiv_id":"2509.07456","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"Applying gradient ascent, LoRA, and distillation unlearning to biased vision models can reduce some subgroup disparities, but the headline accuracy and metric claims are not supported by the paper's own table.","lead":"This paper tests whether machine unlearning, a technique for making models forget specified data, can be used to remove bias from vision models. It applies five unlearning methods to three image datasets and introduces a composite score, Co-BUM, to measure fairness, accuracy, and privacy together.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fairness evaluation is confounded: D_f and the DP/EO partitions share the same subgroup axis, so reported gains may be mechanical forgetting rather than debiasing; Table 2's large accuracy drops make this concrete.","rationale":"Reader's weakest assumption identifies exactly this axis-confounding; I agree. The paper has useful components—GradCAM qualitative analysis, MIA scores, and a systematic comparison of five unlearning methods—but none of these resolves the confound, because the central fairness metric is computed on the same partition used to build D_f. There is also a separate, easily verified contradiction: Table 2 shows the largest DP gains come with large test-accuracy drops, which contradicts the abstract's 'minimal accuracy loss.' That strengthens the case that the fairness gains are not clean debiasing. However, the most fundamental issue remains the evaluation design: it cannot distinguish 'fairer model' from 'model that fails on the forget group.' The proposed held-out-axis test would settle this. If the issue lands, the central claim—machine unlearning as a practical debiasing framework—is unsupported, so the reader's REJECT verdict should stand unchanged.","tokens_in":10516,"tokens_out":12014,"duration_ms":136785,"concrete_test":"Run a held-out-axis evaluation: keep the unlearning procedures exactly as described (D_f defined by pose/patch/gender-smiling), but compute DP and EO on test partitions defined by an independent, previously unused spurious attribute—e.g., background type or bounding-box aspect ratio for CUB; a second synthetic patch location/color for CIFAR; age or attractiveness for CelebA. If these held-out-axis fairness metrics do not improve (or worsen) while the original-axis metrics improve, the reported gains are axis-specific/mechanical and the central claim fails. A useful control is to compare against a 'label-shuffle' unlearning on D_f: if random-label forgetting reproduces the original DP improvement, the improvement is mechanical.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim treats DP/EO improvements as evidence of genuine debiasing. That requires the evaluation axis to be able to distinguish real equity gains from the mechanical effect of degrading predictions on the forget set. In all three settings the same subgroup axis defines both D_f and the fairness partitions: CUB distant-pose bin (§3.1), CIFAR bird+patch samples (§3.2), and the gender-smiling group in CelebA (§3.3). Gradient Ascent and similar methods are explicitly designed to maximize loss on D_f; making the model worse on those samples will tend to compress positive-rate gaps, exactly what DP records. The paper's own §4 warns that 'fairness gains may be superficial rather than substantive' and uses the RA-TA gap as a diagnostic of 'indiscriminate degradation'—by that diagnostic, CelebA GA (RA=58.31 vs TA=50.90) is indiscriminate degradation, yet it is the source of the headline 97.37% DP improvement. Therefore the reported improvements do not yet establish that unlearning debiases the model; they may only show that targeted forgetting changes predictions on the forgotten subgroup.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 'bias-aware machine unlearning' as a post-hoc debiasing tool for vision models. It evaluates Gradient Ascent, LoRA, SCRUB-style teacher-student distillation, Fast Model Debiasing (FMD), and hard unlearning on three datasets: CUB-200-2011 (pose bias), CIFAR-10 (synthetic patch bias), and CelebA (gender-smile correlation). The authors report large demographic-parity (DP) improvements, claim these are achieved with 'minimal accuracy loss,' and introduce a composite metric, Co-BUM, to jointly score utility, fairness, quality, privacy, and efficiency. The central claim is that unlearning can selectively remove spurious correlations and improve fairness without retraining, making it a practical alternative to conventional debiasing pipelines.","tokens_in":10873,"tokens_out":4196,"duration_ms":49265,"significance":"If the empirical claims were reliable, the paper would offer a useful practical contribution: a post-hoc, retraining-free debiasing toolkit with a unified evaluation metric. The study has positive features: it benchmarks multiple unlearning methods across three distinct bias types, includes privacy (MIA) and efficiency metrics, and provides qualitative GradCAM inspection. The paper also explicitly acknowledges in §4 that 'fairness gains may be superficial rather than substantive,' which shows a degree of self-awareness. However, the headline result is undermined by two load-bearing problems: (i) the 'minimal accuracy loss' claim is directly contradicted by the paper's own Table 2, and (ii) the fairness evaluation is confounded because D_f and the DP/EO partitions share the same subgroup axis, so the reported DP improvements may be a mechanical consequence of degrading forget-set predictions. In addition, the Co-BUM metric is not reproducible as defined, because the normalization N_X is never specified and the displayed formula is malformed. These issues prevent the current manuscript from supporting its central claims.","major_comments":[{"comment":"The claim of 'minimal accuracy loss' is contradicted by the reported results. In Table 2, CelebA Gradient Ascent drops test accuracy from 94.20 to 50.90 (an absolute drop of 43.3 points), CUB LoRA drops from 78.69 to 57.47, and CUB Gradient Ascent drops to 64.41. Even methods with the best fairness gains show substantial utility loss. Since 'minimal accuracy loss' is a central selling point of the paper, this contradiction is load-bearing and the claim must be retracted or substantially qualified.","section":"Abstract, Table 2"},{"comment":"Co-BUM, the proposed unified metric, is not reproducible as defined. Figure 2 defines Co-BUM in terms of a normalization N_X, but N_X is never defined in the paper or in Table 1, which only lists the weights α_i and scaling factor κ. Moreover, the displayed formula is garbled: 'Co-BUM := κ (Σ α_i) (Σ α_i / i)' is not a recognizable weighted harmonic mean, and the division and summation indices are ambiguous. Without a precise definition of N_X and the overall formula, the Co-BUM scores in Table 2 cannot be reproduced or compared across methods.","section":"Figure 2, Table 1"},{"comment":"The fairness evaluation is confounded with the unlearning intervention. In each setting, D_f is defined as the biased subgroup (distant-pose birds, bird-with-patch, female-smiling faces), and DP/EO are computed on test partitions of the same subgroup axis. Unlearning methods such as Gradient Ascent are explicitly designed to maximize loss on D_f, so they directly change predictions on that group; the DP improvement may be a mechanical consequence of degrading forget-set performance rather than a genuine reduction in bias. The paper's own diagnostic in §4—'large [RA–TA] gaps indicate indiscriminate degradation'—applies to the CelebA Gradient Ascent result: retain accuracy drops from 96.13 to 58.31, a massive collapse, yet this method produces the headline 97.37% DP improvement. The reported gains therefore do not establish that unlearning debiases the model.","section":"§3.1–3.3, §4"},{"comment":"No experiment separates 'fairness improvement' from 'forgetting the subgroup.' To support the central claim, the authors would need an independent evaluation target: for example, measuring fairness on an attribute not used to construct D_f, or comparing against a control that randomly degrades predictions on D_f to show that the DP gain exceeds the mechanical effect. Without such a control, the causal claim that unlearning removes spurious correlations rather than simply damaging the forget set is unsupported.","section":"§3–§4 (overall design)"}],"minor_comments":[{"comment":"Several rows in Table 2 are typeset with concatenated values (e.g., the CUB Gradient Ascent row appears as 'Gradient Ascent3774.67 64.41 93.75 ...'), making the table difficult to read and potentially ambiguous. Please reformat.","section":"Table 2"},{"comment":"The statement that FMD is 'rendered irrelevant' due to lack of counterfactuals is terse; a more explicit rationale would help readers understand why the comparison is omitted in this setting.","section":"§3.3"},{"comment":"Some references are incomplete or lack venue details (e.g., [27] has no conference/proceedings information). Please ensure all citations are fully specified.","section":"References"},{"comment":"The phrase 'minimal accuracy loss' should be replaced with language that accurately reflects the utility-fairness trade-offs shown in Table 2, or the accuracy figures should be reported with confidence intervals to support the claim.","section":"Abstract"},{"comment":"The caption attributes the CIFAR-10 debiasing to 'the SCRUB framework,' while Section 2 describes a teacher-student method 'inspired by SCRUB.' Please clarify whether SCRUB is used exactly as in the original paper or as an adapted variant.","section":"Figure 1 caption"}],"recommendation":"reject","confidential_remarks":"The paper's empirical core is not reliable: the utility claim is contradicted by the authors' own Table 2, and the fairness evaluation is confounded with the unlearning intervention. The Co-BUM metric is non-reproducible as presented. These are not local presentation issues; they bear directly on the central claims. A revision would require redesigning the evaluation, adding control conditions, and likely changing the headline conclusions, which goes beyond a standard major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper asks the right question—can post-hoc unlearning debias a deployed vision model without retraining—and the comparative setup across three bias types is a reasonable way to approach it. Credit where due: the authors test five unlearning methods on three distinct bias scenarios, include qualitative GradCAM evidence, and in Section 4 they even warn that 'fairness gains may be superficial rather than substantive' and use the RA-TA gap as a diagnostic for indiscriminate degradation. That shows they are thinking about the right pitfalls.\n\nBut the central claims do not survive contact with the data. The abstract promises 'minimal accuracy loss,' yet Table 2 shows the biggest fairness gains coming with large accuracy drops: CelebA Gradient Ascent drops test accuracy from 94.20 to 50.90, and CUB Gradient Ascent from 78.69 to 64.41 (with SCRUB at 45.72). By the paper's own RA-TA diagnostic, the CelebA GA result (RA=58.31 vs TA=50.90) is close to indiscriminate degradation, yet that is the source of the 97.37% DP improvement. So the headline fairness improvements are not evidence of genuine debiasing; they are, at least partly, the mechanical result of making the model worse on the forget set.\n\nThe evaluation design compounds this. The forget set D_f and the DP/EO partitions are defined on the same subgroup axis in all three datasets. Unlearning on D_f directly changes predictions for that group, so any measured gain in demographic parity is confounded with the forgetting objective itself. A proper fairness evaluation needs an independent axis or at least a held-out subgroup not used to define D_f.\n\nThere is also a reproducibility problem: the Co-BUM metric in Figure 2 uses a normalization N_X that is never defined. Without that, the headline average of 0.62 cannot be recalculated or trusted. The formula itself is also typeset in a way that is hard to parse.\n\nWhat is genuinely novel is thin. FMD [7] and Fair Machine Unlearning [20] already frame unlearning as a debiasing tool; this paper is a broad benchmark of known methods with a modified composite metric. That incremental contribution could still be useful if the metric were precise and the evaluation deconfounded, so I would not desk-reject it.\n\nFor a reader: this is a cautionary example of how easy it is to overclaim fairness gains from a confounded evaluation. It deserves referee time, but the authors need major revisions: define N_X, separate D_f from the fairness partitions, report accuracy tradeoffs honestly, and tone down the abstract. I would not cite it in its current form.","headline":"Useful benchmark idea undermined by an undefined metric and an evaluation that shares the same subgroup axis with the forget set, so the headline fairness gains may be mechanical.","tokens_in":11338,"tokens_out":3426,"would_cite":false,"duration_ms":39613,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Post-hoc machine unlearning can cut vision-model subgroup disparities by up to 97%, the paper argues, with only minimal accuracy loss.","keywords":["machine unlearning","bias mitigation","spurious correlations","demographic parity","equalized odds","LoRA","gradient ascent","model fairness"],"falsifier":"Take the unlearned CelebA and CUB models and evaluate them on a held-out test set where the spurious attribute is decorrelated from the label but sampled independently of the forget-set partition (e.g., newly collected bird poses or a gender-balanced smiling test set); if the demographic-parity and equalized-odds improvements disappear or accuracy collapses, the central claim that spurious correlations were removed would be refuted.","tokens_in":10442,"feed_emoji":"⚖️","tokens_out":4997,"duration_ms":52396,"temperature":0.7,"texified_at":"2026-08-05T20:26:55.542233+00:00","pith_summary":"The paper tries to establish that machine unlearning—methods originally built to delete data for privacy—can be repurposed as a post-hoc bias mitigation tool for deployed vision models. It defines a biased subset of the training data, applies five unlearning strategies to remove that subset's influence, and measures fairness on the affected subgroups. Across three benchmark settings—bird pose, synthetic patch shortcut, and gender-smile correlation—it reports large demographic-parity improvements (up to 94.86%, 30.28%, and 97.37%) with test accuracy largely preserved. The paper also proposes a single composite score, Co-BUM, that balances unlearning quality, utility, fairness, privacy, and efficiency so that method choices can be compared on one scale. If correct, the work implies that fairness fixes need not require full retraining or data pipeline redesign.","texify_model":"deepseek-v4-flash","texify_usage":{"total_tokens":4740,"prompt_tokens":731,"completion_tokens":4009,"prompt_tokens_details":{"cached_tokens":0},"prompt_cache_hit_tokens":0,"prompt_cache_miss_tokens":731,"completion_tokens_details":{"reasoning_tokens":3370}},"feed_headline":"Forgetting biased samples cuts vision bias by up to 97%","feed_subtitle":"Post-hoc unlearning improves demographic parity on three benchmarks without full retraining, the paper reports.","key_machinery":"The load-bearing object is the bias-aware unlearning formulation: identify a biased subset $D_b$ of the training data, then update parameters so the model approximates one retrained on $D_{\\text{train}} \\setminus D_b$. The update mechanisms are gradient ascent on the forget-set loss with retain-set regularization, LoRA low-rank adapters maximizing forget loss while minimizing retain loss, a teacher-student distillation objective that aligns on retain data and diverges on forget data, and a counterfactual influence-function update. The evaluation machinery is Co-BUM, a weighted harmonic mean that normalizes utility, fairness (Demographic Parity and Equalized Odds), unlearning quality, membership-inference privacy, a","core_discovery":"The central claim is that selectively forgetting a hand-chosen biased subgroup, rather than retraining, can remove spurious correlations and reduce group disparities in vision classifiers. The paper formalizes bias-aware unlearning as updating model parameters so the resulting function approximates a model retrained on the data minus the biased subset, and tests gradient ascent, LoRA fine-tuning, teacher-student distillation, counterfactual debiasing, and exact retraining. The result pattern is context-dependent: boundary-pushing gradient ascent works best for diffuse pose bias, low-rank adapters work best for a localized patch artifact, and aggressive forgetting nearly eliminates demographi","pith_inferences":["Editorial extension: because the forget set and the fairness evaluation set are the same hand-defined subgroup, the reported DP gains may partly reflect the mechanical effect of changing predictions on exactly those examples; testing on an independent spurious-correlation partition would separate genuine debiasing from memorization of the intervention.","Editorial extension: the promise of avoiding full retraining would be strengthened by comparing against standard debiasing baselines such as reweighting, adversarial training, and data augmentation under matched compute budgets; the paper does not include that comparison.","Editorial extension: a natural extension is sequential or joint unlearning of multiple correlated biases, since the paper treats each bias in isolation.","Editorial extension: the Co-BUM normalization against a 'gold' retrained model requires access to a retrained reference model, which may limit its use exactly in the no-retraining deployment scenario the paper motivates."],"forward_implications":["Deployed vision models that have learned spurious correlations can be corrected in minutes-to-hours on a single GPU rather than retrained from scratch.","The best unlearning method depends on bias type: boundary-pushing methods for distributed pose bias, small adapter updates for localized shortcuts, and aggressive forgetting for entrenched attribute correlations.","Fairness gains can co-occur with privacy gains: several methods lower membership-inference attack success as they lower demographic parity gaps.","Demographic parity and equalized odds move apart after unlearning, so reporting only one fairness metric can overstate progress.","Co-BUM provides a single number for comparing unlearning strategies across utility, fairness, quality, and privacy, aiding post-deployment correction choices."],"supporting_citations":[{"why":"Supplies the counterfactual debiasing framework the paper benchmarks as FMD.","marker":"[7]"},{"why":"Supplies the teacher-student distillation objective used as SCRUB.","marker":"[16]"},{"why":"Supplies the gradient ascent update rule with retain-set regularization.","marker":"[23]"},{"why":"Formalizes the target of approximating a model retrained on data without the biased subset.","marker":"[20]"},{"why":"Supplies the multi-objective evaluation setup the paper extends into Co-BUM.","marker":"[18]"},{"why":"Defines Demographic Parity used for fairness measurement.","marker":"[5]"},{"why":"Defines Equalized Odds used for fairness measurement.","marker":"[12]"},{"why":"Supplies the CelebA dataset and the gender-smile bias scenario.","marker":"[17]"},{"why":"Supplies the shortcut-learning framing that motivates separating semantic from spurious features.","marker":"[10]"},{"why":"Supplies the ResNet-18 backbone used across all experiments.","marker":"[13]"}],"fun_headline_variants":["Forgetting biased data cuts vision bias by 97% without retraining","Machine unlearning slashes vision bias up to 97%","Selective forgetting removes spurious correlations in vision models","Unlearning biased samples boosts fairness up to 97%","Forget bias: post-hoc unlearning improves vision fairness"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The validity of the headline gains rests on the assumption that unlearning on the hand-chosen biased subset—distant-pose birds, patched birds, male-smiling faces—is the right intervention axis, and that measuring fairness on exactly those subgroups reflects genuine debiasing rather than a mechanical prediction shift.","fun_headline_variants_meta":{"raw":{"variants":["Forgetting biased data cuts vision bias by 97% without retraining","Machine unlearning slashes vision bias up to 97%","Selective forgetting removes spurious correlations in vision models","Unlearning biased samples boosts fairness up to 97%","Forget bias: post-hoc unlearning improves vision fairness"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000207,"raw_usage":{"total_tokens":1256,"prompt_tokens":779,"completion_tokens":477,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":393}},"tokens_in":523,"tokens_out":477,"duration_ms":5374,"temperature":1.0,"reasoning_tokens":393,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T22:09:59.698614+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the unlearned CelebA and CUB models and evaluate them on a held-out test set where the spurious attribute is decorrelated from the label but sampled independently of the forget-set partition (e.g., newly collected bird poses or a gender-balanced smiling test set); if the demographic-parity and equalized-odds improvements disappear or accuracy collapses, the central claim that spurious correlations were removed would be refuted.","supporting_citations":[{"cited_title":"Fast model debias with machine un- learning","cited_arxiv_id":null,"evidence_quote":"Supplies the counterfactual debiasing framework the paper benchmarks as FMD."},{"cited_title":"Towards unbounded machine unlearning","cited_arxiv_id":null,"evidence_quote":"Supplies the teacher-student distillation objective used as SCRUB."},{"cited_title":"Unrolling SGD: Understanding Factors Influencing Machine Unlearning","cited_arxiv_id":"2109.13398","evidence_quote":"Supplies the gradient ascent update rule with retain-set regularization."},{"cited_title":"Fair machine unlearning: Data removal while mitigating disparities","cited_arxiv_id":null,"evidence_quote":"Formalizes the target of approximating a model retrained on data without the biased subset."},{"cited_title":"Breaking the trilemma of privacy, utility, and efficiency via controllable machine unlearning","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-objective evaluation setup the paper extends into Co-BUM."},{"cited_title":"Building classifiers with independency constraints","cited_arxiv_id":null,"evidence_quote":"Defines Demographic Parity used for fairness measurement."},{"cited_title":"Equality of opportunity in supervised learning","cited_arxiv_id":null,"evidence_quote":"Defines Equalized Odds used for fairness measurement."},{"cited_title":"Deep learning face attributes in the wild","cited_arxiv_id":null,"evidence_quote":"Supplies the CelebA dataset and the gender-smile bias scenario."},{"cited_title":"Shortcut learning in deep neural networks","cited_arxiv_id":null,"evidence_quote":"Supplies the shortcut-learning framing that motivates separating semantic from spurious features."}],"review_version":1}