{"id":"b569d540-8e5e-4977-8d94-21b4e4ef6abe","arxiv_id":"2411.11101","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Bias mitigation algorithms in ML are highly sensitive to hyperparameters, and most can achieve comparable fairness-accuracy tradeoffs once hyperparameter optimization is allowed, so single-setting benchmarks can mislead.","lead":"This paper compares seven machine-learning fairness methods under many different training settings and shows their rankings change with the settings. It argues that fixed-setup benchmarks can make one method look best when most methods are actually competitive after hyperparameter tuning.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pareto-front comparability is assessed visually, with no quantitative 'comparable' criterion and no accuracy floor, so degenerate low-utility points can make algorithms look competitive; grid sparsity in Appendix A compounds this.","rationale":"The paper makes a useful and mostly credible empirical point: single-setting benchmarks are fragile, and several algorithms can be tuned to similar regions of the fairness-accuracy plane. The authors build on an existing open benchmark (FFB), provide a large number of scatter plots across datasets and metrics, and add independent evidence on runtime and multiplicity that does not depend on the comparability claim. I found no internal inconsistency or circular reasoning. The soft spot is that the headline conclusion is drawn from visual overlap of Pareto fronts rather than from a quantitative notion of 'comparable.' The absence of an accuracy floor is particularly concerning because a Pareto front always includes extreme tradeoff points; any algorithm that can sacrifice accuracy can put a point on the front. The reader's worry about grid coverage is related but secondary: even with perfect coverage of hyperparameter space, the conclusion would still be underdetermined without a criterion for comparability and a guard against degenerate models. My recommendation is therefore to keep the reader's CONDITIONAL verdict, with the additional explicit condition that comparability be demonstrated quantitatively under a utility floor. If the authors supply such an analysis and it confirms the visual impression, the central claim would be substantially stronger; if not, the paper should soften the abstract to 'can achieve comparable or near-comparable performance on some evaluated settings' rather than the current generalization.","tokens_in":945,"tokens_out":1722,"duration_ms":63069,"concrete_test":"Re-analyze the collected runs (or rerun the grid) as follows: for each dataset and fairness metric, restrict to models whose accuracy is within X% (e.g., 5%) of the ERM accuracy under the same hyperparameter setting, or at least within a fixed absolute margin. Then compute a quantitative comparability score for each algorithm, for example the maximum fairness achieved at each accuracy level, the area under the Pareto front, or the minimum accuracy loss relative to the best front at each fairness level. Declare algorithms 'comparable' only if their score lies within a pre-specified tolerance, such as never being more than 1-2% accuracy below the best front at any fairness level. If fewer than 'most' algorithms survive this test, the abstract's claim needs qualification; if most survive, the claim is supported and the grid-coverage worry is less central.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim ('most mitigation techniques can achieve comparable performance, given the freedom to perform hyperparameter optimization') is supported primarily by visual overlap of per-algorithm Pareto fronts in Figure 4 and Appendix D. Two load-bearing conditions are never made explicit. First, 'comparable' has no quantitative definition: no threshold on accuracy loss at a given fairness level, no Pareto-dominance gap, and no area-under-front comparison. Second, because points are only required to lie on the fairness-accuracy Pareto front, an algorithm can appear 'competitive' by contributing a degenerate high-fairness/low-accuracy corner; such models would not be deployed, contrary to Section 4.1's claim that the selected models are 'the models that would be deployed.' Without an accuracy floor (e.g., within a small margin of ERM accuracy) and a quantitative comparability criterion, the conclusion that most algorithms are interchangeable on fairness-utility tradeoffs is not falsifiable from the presented plots. The small grid (7 hyperparameter settings, 3 control values, unreported number of random seeds, Appendix A) aggravates this: selecting the envelope of a small grid can make algorithms look comparable by chance, especially if seed counts differ across algorithms. The paper's independent support (runtime and multiplicity tables) is valuable, but those secondary comparisons do not rescue the headline comparability claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the sensitivity of in-processing fairness mitigation algorithms to hyperparameter choices, random seeds, and pipeline settings. Using the FFB benchmark's seven tabular datasets and seven mitigation algorithms, the authors vary batch size, learning rate, architecture, control parameter, and seeds, and plot fairness-accuracy tradeoffs under demographic parity, equalized odds, and equal opportunity. They argue that no algorithm dominates across settings, that FFB's single-setting and cross-dataset aggregation hides dataset-specific trends, and that after hyperparameter search most algorithms achieve comparable Pareto-front performance. They then propose selecting algorithms by additional criteria such as runtime, theoretical guarantees, and predictive multiplicity, and support this with runtime and ambiguity tables.","tokens_in":11423,"tokens_out":4904,"duration_ms":51625,"significance":"If substantiated, the central claim that 'most bias mitigation techniques can achieve comparable performance, given the freedom to perform hyperparameter optimization' would challenge the validity of single-setting fairness benchmarks such as FFB and would redirect algorithm selection toward deployment-related criteria. The paper's strengths are its breadth (seven datasets, seven algorithms, three fairness metrics), its use of an external benchmark rather than self-constructed settings, and its secondary empirical contributions: runtime measurements and multiplicity/ambiguity tables that offer independent grounds for algorithm choice. The paper is also transparent in Section 5 about its scope (in-processing methods, training-time hyperparameters only). However, the headline comparability claim currently rests on visual inspection of Pareto fronts with no quantitative definition of 'comparable', no accuracy floor, no statistical support, and a relatively small and under-specified hyperparameter grid. These gaps are load-bearing for the main conclusion.","major_comments":[{"comment":"The central claim that 'most mitigation techniques can achieve comparable performance' is not attached to any quantitative definition of 'comparable.' The evidence is visual overlap of per-algorithm Pareto fronts, but any point on a Pareto front, including a degenerate high-fairness/low-accuracy point that would not be deployed, makes its algorithm appear 'competitive.' The text in Section 4.1 says these are 'the models that would be deployed,' yet without an accuracy floor (e.g., within a stated margin of ERM or of the best algorithm's accuracy) the displayed fronts do not establish that. Please define a comparability criterion, such as maximum allowed accuracy loss at a target fairness level, Pareto dominance gap, or hypervolume under a utility floor, and apply it to the reported results.","section":"Section 4.1, Figures 4, 15, 16"},{"comment":"The hyperparameter search underlying the Pareto-front claim consists of seven hyperparameter settings and three control values, with the number of random seeds left unspecified. No coverage analysis, convergence check, or comparison across seeds is reported. With as few as 21 runs per algorithm per dataset, the Pareto front is an envelope of a small grid rather than a characterization of what the algorithms can achieve under hyperparameter optimization; a different or denser grid could change which algorithms appear competitive. Please report the number of seeds, use identical seeds across algorithms and datasets, and provide a sensitivity analysis of the Pareto fronts to the grid, or use a larger random search.","section":"Appendix A, Tables 3 and 4, Section 4.1"},{"comment":"No error bars, confidence intervals, or statistical tests are provided, so the claimed 'significant variance' and 'comparable performance' are not distinguished from run-to-run noise. For instance, Table 2 reports large multiplicity differences (e.g., AdvDebias 0.87 vs. ERM 0.55 on German) without any measure of variability across seeds, and Figure 3 presents dataset-level tradeoffs as point clouds with no summary statistics. Please report distributions over seeds and control values (mean, standard deviation, or min/max) and apply a statistical test or effect-size measure to the pairwise algorithm comparisons that underlie the headline conclusion.","section":"Figures 2-4, Tables 1-2"},{"comment":"The dataset-specific explanations, such as HSIC's weakness on small-batch datasets and LAFTR's sensitivity to one-hot-encoded features, are post hoc and not tested. These explanations are used to argue that FFB-style cross-dataset aggregation obscures trends, so they need at least a controlled comparison, e.g., varying batch size while holding other hyperparameters fixed, or varying the feature encoding while holding the algorithm fixed. Without such tests, these claims remain speculative and should be labeled as hypotheses rather than findings.","section":"Section 3.3"}],"minor_comments":[{"comment":"The table lists 'DiffEOdd' twice and omits DiffEOpp; the second row should presumably be DiffEOpp.","section":"Table 3"},{"comment":"The number of random seeds used for each setting is never stated; please include this information in the table or text.","section":"Appendix A.3"},{"comment":"The captions use lowercase 'pareto front'; please use 'Pareto front' consistently and define the term at first use in Section 4.1.","section":"Figure 4 and Appendix D"},{"comment":"The 'Baharlouei et al.' reference lacks a year and publication venue; please complete it.","section":"References"},{"comment":"The paper borrows the FFB codebase but does not release its own modified code or configuration files; a reproducibility statement listing exact changes (e.g., which hyperparameters were added, how seeds were handled) would strengthen the paper.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"This is a workshop-style empirical study whose central claim needs substantial additional quantification before it meets the standard for a journal. I do not see circularity in the use of the external benchmark; the main issue is an under-specified and partly unfalsifiable comparability claim. Code release and explicit seed counts would also help, since the paper's contribution is empirical. The runtime and multiplicity tables are the most solid parts of the paper and could be foregrounded if the Pareto-front claim is made more rigorous."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read.\n\nThe paper's real contribution is a concrete demonstration on the FFB benchmark that most of the seven in-processing bias mitigation algorithms land on similar fairness-accuracy Pareto fronts once you search over hyperparameters, plus runtime and ambiguity comparisons that give practitioners something beyond tradeoff curves. That is useful. It reframes algorithm selection around stability, runtime, and multiplicity rather than single-setting leaderboard rank. The related-work handling is fair: it cites the variance and multiverse literature (Black, Friedler, Simson) and correctly notes that FFB aggregation can hide dataset-specific behavior.\n\nIt does several things well. It reuses an external benchmark and extends it with multiple hyperparameter settings (batch size, learning rate, layer width) and three control values per algorithm. The scatter plots across datasets and settings do show that the 'best' algorithm changes with context. The runtime table and ambiguity scores are a real addition; the finding that AdvDebias consistently has high ambiguity and PRemover swings between extremes is interesting and not just a restatement of the tradeoff point.\n\nSoft spots, in proportion. The headline claim—'most bias mitigation techniques can achieve comparable performance, given the freedom to perform hyperparameter optimization'—is supported by visual inspection of Pareto fronts. 'Comparable' never gets a quantitative definition. There is no accuracy floor, so an algorithm can appear competitive by contributing a degenerate high-fairness/low-accuracy corner that no one would deploy, undercutting the claim in Section 4.1 that these are 'the models that would be deployed.' The grid is small: seven hyperparameter settings, three control values, and the number of random seeds is not reported. Selecting the envelope of a small grid can make algorithms look similar by chance, especially if seed counts differ. No error bars, no statistical test, and no released code. These are real limitations, but they soften the claim rather than sink it. The paper is honest about scope: it covers only in-processing methods and only training-time hyperparameters, and says so in the Discussion.\n\nMy take: the central argument holds up in spirit. The variance is real, context-dependence is real, and the FFB-based demonstration is a reasonable first pass. But the comparability conclusion is not yet falsifiable from the presented figures. For a workshop paper this is acceptable; for a journal it would need an operationalized comparability metric, an accuracy floor, and ideally released code and seeds.\n\nWho this is for: people building fairness benchmarks and practitioners choosing among mitigation algorithms. It deserves a serious referee: the question is important, the experimental design is reasonable, and the runtime/multiplicity data are worth having on record. I would send it to review, with a request for quantitative tightness.","headline":"Useful empirical counterpoint to single-setting fairness leaderboards, but the parity claim rests on visual Pareto overlap rather than a quantitative comparability criterion.","tokens_in":11898,"tokens_out":3427,"would_cite":false,"duration_ms":98717,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that most bias mitigation algorithms can achieve comparable fairness–accuracy tradeoffs once hyperparameters are optimized, so perceived superiority may reflect evaluation setup rather than the algorithm.","keywords":["fairness in machine learning","bias mitigation algorithms","hyperparameter sensitivity","fairness–utility tradeoff","Pareto front","model multiplicity","benchmarking","demographic parity"],"falsifier":"Run the same seven algorithms on the same datasets with a substantially expanded hyperparameter search—for instance, a Bayesian optimization budget with dozens of trials per algorithm—and recompute each algorithm's Pareto front. If some algorithms that look competitive under the paper's grid are consistently dominated, or if algorithms that look weak are pulled onto the front, the claim that most techniques are competitive would fail; a quantitative coverage analysis showing the fronts still move materially after additional search would also falsify it.","tokens_in":10973,"feed_emoji":"⚖️","tokens_out":7143,"duration_ms":64474,"temperature":0.7,"pith_summary":"Bias mitigation algorithms in machine learning are usually benchmarked under one uniform setup, with the assumption that identical training conditions make comparisons fair. This paper argues that those conditions are anything but neutral: fairness scores swing with random seeds, batch size, learning rate, model architecture, and the algorithm's own control parameter. Running seven mitigation algorithms on seven tabular datasets across multiple hyperparameter settings, the authors find no consistently winning method; different algorithms lead under different settings, and some fail to converge while others excel. Filtering each algorithm to the models on the fairness–accuracy Pareto front after a hyperparameter search, they conclude that most mitigation techniques can reach comparable tradeoffs. If true, algorithm rankings in fairness benchmarks can be an artifact of the chosen evaluation setup, and choosing an algorithm should hinge on context—runtime, guarantees, robustness, and arbitrariness—rather than a single tradeoff curve.","feed_headline":"Tuning erases most fairness-algorithm gaps","feed_subtitle":"Across seven datasets, no bias-mitigation method dominates once hyperparameters are optimized; context should decide.","key_machinery":"The load-bearing object is the Pareto front of the fairness–utility tradeoff: the set of models from each algorithm that are not dominated on both demographic parity (or equalized odds or equal opportunity) and accuracy. The authors sweep seven hyperparameter settings per dataset, three control parameter values per algorithm, and multiple random seeds, then keep only the non-dominated models for each algorithm. Comparing these fronts is what turns 'which algorithm is best' into 'which algorithms can reach a given tradeoff at all,' and it is what reveals that multiple methods share nearly the same frontier on most datasets. The same object also carries the negative result: under a single fixed setting, the front collapses to one point per algorithm, and rankings that look stable are actually slices through a much larger and overlapping cloud of achievable outcomes.","core_discovery":"On the paper's own terms, the central discovery is that the fairness–utility ordering of bias mitigation techniques is unstable across the learning pipeline, and that this instability undermines one-dimensional benchmarking. Using the benchmark's seven datasets and seven in-processing algorithms, the authors vary random seeds, control parameters, batch size, learning rate, and architecture, and plot demographic parity and accuracy tradeoffs. They observe that no algorithm dominates: HSIC, which the benchmark singled out as offering the best tradeoff, is matched or beaten by DiffDP and PRemover under other settings, and algorithms like AdvDebias and LAFTR only stand out under large-batch, high-learning-rate regimes where others fail to converge. Across datasets, HSIC's advantage fades on small datasets like COMPAS and German, and LAFTR's weakness appears tied to one-hot feature explosion rather than to the method alone. When the authors form the Pareto front of each algorithm over the searched hyperparameters, several algorithms occupy similar fronts on almost every dataset, which they take as evidence that most mitigation techniques are competitive given freedom to tune. They then show that runtime and prediction multiplicity distinguish algorithms even when tradeoffs do not.","pith_inferences":["If the comparability result generalizes, fairness leaderboards should shift from point rankings to reporting Pareto fronts and variance bands; otherwise they will keep rewarding whichever algorithm happens to fit the chosen defaults.","A natural testable extension is to check whether the same equalizing effect appears under automated hyperparameter optimization with larger budgets and across pre- and post-processing methods, which the paper does not cover.","The multiplicity results suggest a possible tradeoff beyond fairness–utility: algorithms that enforce fairness through representation learning may increase individual-level arbitrariness, which matters for contestability and recourse even when aggregate fairness looks identical.","Regulatory reporting of fairness metrics could be gamed through cherry-picked hyperparameters; the paper's variance results imply that reports should include a range or distribution of fairness scores over training configurations, not a single number."],"forward_implications":["No single bias mitigation algorithm should be declared the best from a single hyperparameter configuration; results should report the spread of fairness scores across settings.","In ranking or regulator-facing evaluations, hyperparameter optimization should be part of the protocol, because restricted comparisons can unfairly favor or penalize algorithms.","Given comparable tradeoffs, selection criteria like training runtime, theoretical fairness guarantees, and prediction multiplicity become decisive; for example, DiffDP is attractive for fast training and PRemover for low ambiguity on some datasets.","Dataset aggregation in benchmarks can hide dataset-specific failures, such as HSIC's dependence on batch size for estimating statistical dependence.","Deployment context—whether retraining is cheap or expensive, whether stability or peak performance matters—should determine how fairness evaluations are conducted."],"supporting_citations":[{"why":"Provides the benchmark datasets, algorithms, and code that the study extends, and supplies the single-setting claim that HSIC is best.","marker":"[Han et al., 2023]"},{"why":"This comparative study is cited as evidence that fairness interventions vary with training setup and train-test splits.","marker":"[Friedler et al., 2019]"},{"why":"Introduces the pipeline-aware fairness agenda and the instability that can create an illusion of fairness in fair ML systems.","marker":"[Black et al., 2023]"},{"why":"Shows that small random-seed changes can flip group fairness outcomes, supporting the paper's claim that single-seed comparisons are unreliable.","marker":"[Black and Fredrikson, 2021]"},{"why":"Documents how randomness in ML training affects group fairness, used to justify the focus on high-variance in-processing methods.","marker":"[Ganesh et al., 2023]"},{"why":"Demonstrates sensitivity of fairness methods to hyperparameter choice, motivating the multi-setting search.","marker":"[Perrone et al., 2021]"},{"why":"Defines the set of competing models and individual arbitrariness used in the multiplicity analysis.","marker":"[Long et al., 2024]"},{"why":"Defines predictive multiplicity and ambiguity, the metric the paper uses to compare arbitrariness across algorithms.","marker":"[Marx et al., 2020]"}],"fun_headline_variants":["Hyperparameter tuning reshuffles fairness rankings","Fairness algorithms tie once tuned","Tuning flips fairness leaderboard","Context, not algorithm, sets fairness","No dominant debiasing method after tuning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the hyperparameter and seed grid searched here is wide enough to reveal each algorithm's true range of fairness–accuracy outcomes, even though the paper does not check whether a broader or different search would change the rankings.","fun_headline_variants_meta":{"raw":{"variants":["Hyperparameter tuning reshuffles fairness rankings","Fairness algorithms tie once tuned","Tuning flips fairness leaderboard","Context, not algorithm, sets fairness","No dominant debiasing method after tuning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000148,"raw_usage":{"total_tokens":1184,"prompt_tokens":938,"completion_tokens":246,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":184}},"tokens_in":554,"tokens_out":246,"duration_ms":2873,"temperature":1.0,"reasoning_tokens":184,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:54:01.557655+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same seven algorithms on the same datasets with a substantially expanded hyperparameter search—for instance, a Bayesian optimization budget with dozens of trials per algorithm—and recompute each algorithm's Pareto front. If some algorithms that look competitive under the paper's grid are consistently dominated, or if algorithms that look weak are pulled onto the front, the claim that most techniques are competitive would fail; a quantitative coverage analysis showing the fronts still move materially after additional search would also falsify it.","supporting_citations":[{"cited_title":"[2023], we redirect the reader to their work and the FFB benchmark code 3 for details on the underlying setup","cited_arxiv_id":null,"evidence_quote":"Provides the benchmark datasets, algorithms, and code that the study extends, and supplies the single-setting claim that HSIC is best."}],"review_version":1}