{"id":"3f7d1bf0-873b-4418-8175-f7b72900231f","arxiv_id":"2501.11597","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper defines ECD as the difference in GEV location parameters between protected groups and claims it reveals that standard bias mitigators harm worst-case fairness in 35% of cases.","lead":"This paper proposes a new fairness metric, extreme counterfactual discrimination (ECD), that uses extreme value theory to measure the worst-case discrimination of machine learning models at the tail of outcome distributions. The authors report that common fairness mitigators can reduce average bias but significantly worsen tail bias in a substantial fraction of cases.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ECD compares GEV locations fitted at group-specific thresholds and sample sizes; the reported mu_u - mu_p conflates threshold choice with genuine tail differences.","rationale":"The paper's contribution is an interesting and timely idea: use extreme value theory to measure discrimination in the tail of ML outcomes, not just the average. The experimental effort is substantial, with four models, nine datasets, several mitigation algorithms, and a replication package, and the paper honestly discusses limitations and threats to validity. Those are real strengths. However, the central metric ECD is defined as the difference between GEV location parameters mu_u - mu_p, and those parameters are fitted after choosing group-specific thresholds so that exactly k_max=50 samples exceed the threshold in each group. With group sample sizes differing by an order of magnitude, this means the same number of exceedances sits at very different quantiles of each group's CD distribution. In any threshold-exceedance EVT fit, the location parameter is tied to the chosen threshold; comparing raw locations across groups is therefore not a well-posed comparison of worst-case behavior. The paper even illustrates this: in Figure 1, white and black have different thresholds (0.12 vs. 0.20). The reported ECD value of 0.13 for the DNN Adult example could be mostly a threshold artifact. This is not just an academic nuance: the empirical claims in RQ3 and RQ4 (35%, 30%, 90%) are aggregate counts derived from these ECD values. If the ECD metric is non-comparable across groups, those headline numbers are not supported. The reader's verdict of REJECT is appropriate. A revised version that redefines ECD at a common exceedance probability or via common-return-period return levels, and that reconciles the abstract/body inconsistency (5% vs. 35%) and the arbitrary 0.05 threshold, could be a valuable contribution. But as stated, the central claim does not hold.","tokens_in":27242,"tokens_out":3709,"duration_ms":40738,"concrete_test":"Re-run the pipeline for all 40 baseline scenarios in Table IV with a threshold rule fixed to a common quantile per group, e.g., the 99th percentile of each group's CD distribution (or the top 1% of samples where sample sizes allow), and refit the tail distribution. Then recompute ECD using a threshold-invariant quantity, such as the difference in 1-in-1,000-interaction return levels or the difference in tail quantiles at a common exceedance probability (e.g., 0.01). Compare the resulting ECD values and the RQ3/RQ4 win rates to Tables IV-VII. If the sign or magnitude of ECD changes materially, or the '35% increase' and 'MiniMax 30%' counts move outside the reported ranges, the empirical claims are artifacts of threshold selection rather than genuine worst-case discrimination.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim rests on ECD = mu_u - mu_p (Definition III.2; Section IV, 'Inferring the tail distributions of counterfactual discrimination'). The location parameters are fit to group-specific thresholds: Section II sets tau=0.12 for white and tau=0.20 for black, each chosen so that only 50 samples exceed the threshold. Because the groups' sample sizes differ by an order of magnitude (e.g., 25,658 white vs. 3,076 black in the running example; #N ranges from roughly 0.1k to 25.7k in Table IV), the same k=50 exceedances correspond to very different tail quantiles (approximately the 0.2% vs. the 1.6% extremes). In threshold-exceedance models, the location parameter mu depends on the threshold itself: for the same underlying tail, a higher threshold yields a larger location. Thus ECD = mu_u - mu_p can differ across groups simply because of threshold and sample-size choices, with no adjustment for the exceedance probability being compared. The paper never standardizes thresholds or compares return levels at a common exceedance probability, so the headline empirical statements (average-based mitigators increase worst-case discrimination in 35% of cases; MiniMax increases it in 30%; ECD-Fair improves the tail in 90%) are not interpretable as stated. Since ECD is the dependent variable in RQ2-RQ4, this comparability flaw is load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a fairness criterion called extreme counterfactual discrimination (ECD), defined as the difference between the GEV location parameters mu_u - mu_p fitted to the upper tail of each protected group's counterfactual discrimination (CD) distribution, and argues that this quantifies worst-case discrimination. The authors present a randomized test-case generation algorithm that combines generative AI (CTGAN, TVAE) with a coefficient-of-variation based exponentiality test, and report experiments over 9 datasets and 4 model classes. Their headline claims are that EVT fits the tail distribution in 95% of cases, that average-based bias mitigators increase worst-case discrimination in 35% of cases, that MiniMax-Fairness increases it in 30% of cases, and that their proposed ECD-Fair mitigator improves tail fairness in 90% of cases without degrading average-case fairness.","tokens_in":27533,"tokens_out":5835,"duration_ms":56597,"significance":"The paper addresses a real and underexplored problem: fairness at the extreme tail of outcome distributions, rather than only on average. If the ECD metric were valid, the paper would be a useful addition to fairness testing, providing both an evaluation lens and a mitigation objective. The authors ship a replication package, use standard EVT software, and compare against CVaR and Minimax-Fairness, which are relevant baselines. However, as presented, the metric's definition and fitting procedure make the central quantitative claims (the 35%, 30%, and 90% figures, and the 95% EVT fit rate) not supportable without substantial revision.","major_comments":[{"comment":"The ECD metric is defined as mu_u - mu_p, where mu is the GEV location parameter fitted to each group's threshold exceedances. Because the threshold tau is chosen separately per group so that only k_max = 50 samples exceed it (Section II; Table IV), and the group sample sizes differ by up to two orders of magnitude (#N ranges from roughly 0.1k to 25.7k in Table IV), the same exceedance count corresponds to very different tail quantiles (e.g., roughly the 0.2% vs 1.6% extremes for the white vs black groups in the running example). The location parameter in a threshold-exceedance model depends on the threshold itself, so ECD = mu_u - mu_p conflates threshold and sample-size differences with genuine tail differences. The paper never standardizes thresholds or compares return levels at a common exceedance probability; therefore the headline empirical findings (RQ2, RQ3, RQ4) are not interpretable as stated.","section":"Definition III.2 and Section IV, 'Inferring the tail distributions of counterfactual discrimination'"},{"comment":"The claim that EVT fits in 95% of cases is partly built into the method: Algorithm 1 regenerates samples until the coefficient-of-variation test passes or a timeout occurs, and cases that do not pass are later reported as 'NV' in Tables V and VI. The CV test selects for exponential or light tails, so the 95% figure measures the success of the sampling loop under a permissive selection criterion rather than an independent validation of the GEV fit. The paper should report the number of timeouts, the number of NV exclusions, and evaluate the fitted distributions on held-out tail data (e.g., Q-Q plots and goodness-of-fit tests for all scenarios, not only the selected ones).","section":"Algorithm 1 and Section V-B (Answer RQ2)"},{"comment":"The statistical model is ambiguous. The text states that 'we use the threshold approach where extreme events that exceed some high threshold u are extreme values' and then says 'we fit the GEV distribution,' but the threshold approach for exceedances yields a generalized Pareto distribution, not a GEV. Moreover, the reported 'location' mu values (e.g., 0.15 and 0.28 for white and black in Section II) are larger than the thresholds tau (0.12 and 0.20), which is not the usual parameterization of a threshold-exceedance model. The authors must specify exactly which distribution is fitted, to what data (block maxima of exceedances or all exceedances), and how the location parameter is defined in that fit.","section":"Section IV and Table IV"},{"comment":"The statement 'We classify any ECD difference exceeding 0.05 as discrimination' is used to derive the 48% detection rate and the comparisons in RQ2-RQ4, but no justification, statistical test, or confidence interval is provided for this 0.05 cutoff. The abstract and the RQ answers repeatedly use the word 'significantly' without a corresponding hypothesis test. The paper should report the uncertainty in ECD (e.g., bootstrap or profile-likelihood intervals) and state how the 0.05 threshold was chosen.","section":"Section V-B, paragraph classifying ECD differences"},{"comment":"Several fitted shape parameters are degenerate or implausible: for example, Credit male (DNN) has xi = -141.17, Students male (DNN) has xi = -347.33, and several LR/SVM rows report xi near -1 with mu near 0 and sigma near 0 (e.g., Census White LR). For a bounded outcome (CD is in [-1,1]), such extreme negative shape values imply the fitted distribution has an upper endpoint essentially at the observed maximum, and the reported standard errors of 0.00 (epsilon) are not credible. These cases are nonetheless included in the ECD comparisons; the paper should either exclude or specially treat degenerate fits and report the sensitivity of the RQ3/RQ4 results to their inclusion.","section":"Table IV"}],"minor_comments":[{"comment":"The abstract says average-based mitigators 'increase the worst-case discrimination significantly in 5% of cases,' while Answer RQ3 and the introduction state 35%; the correct number should be reconciled.","section":"Abstract vs. Section V-C"},{"comment":"The text 'TA VE' in the answer to RQ1 is a typo and should read 'TVAE.'","section":"Answer RQ1"},{"comment":"The phrase 'we initially set the threshold of extreme values to Mk max' is unclear; it should say that the threshold is set to the k_max-th largest observed CD value.","section":"Section IV, 'Inferring the tail distributions of counterfactual discrimination'"},{"comment":"The table lists protected attributes for each dataset, but the running example in Section II uses Adult with Race, while the table lists both Sex and Race for Adult; the text should clarify which protected attribute is used in each reported experiment.","section":"Table II"},{"comment":"The figures would benefit from a legend or explicit color labels for the threshold lines and group densities; the captions currently refer to 'red lines' and 'left/mid-left' without a visual legend.","section":"Figures 1 and 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the topic is timely. The main concern is the comparability of ECD across groups with different thresholds and sample sizes, which affects the core empirical claims. I do not see the 95% fit rate as a deliberate circularity, but the procedure does select on the test outcome, so the claim needs repositioning. The authors should be asked to re-run their analysis with a common exceedance probability or with return levels, and to report degenerate fit cases explicitly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper has a genuinely new idea—using extreme value theory to audit worst-case counterfactual discrimination—and a large, mostly careful empirical study. But the headline metric, ECD = µ_u − µ_p, is not comparable across groups as defined, and that flaw is load-bearing. The stress-test note is right.\n\nWhat is new and good: the ECD criterion itself, the use of generative models (CTGAN, TVAE) to produce tail samples, and the empirical observation that average-based mitigators can worsen the tail in a non-trivial share of cases. The paper also ships a replication package, runs four ML models over nine datasets, and includes a limitations section that names real threats (lack of ground truth, protected-attribute dependence). The related-work coverage is fair and mostly on-point.\n\nThe soft spots, in order of severity. First, the comparability flaw: for each group, the threshold is chosen so that only k_max=50 samples exceed it. Group sample sizes differ by an order of magnitude (e.g., 25,658 white vs 3,076 black for Census DNN), so the 50 exceedances sit at very different quantiles. In threshold-exceedance models the location parameter depends on the threshold itself, so µ_u − µ_p conflates genuine tail differences with threshold and sample-size choices. The paper never standardizes to a common exceedance probability or compares return levels at a common quantile. Second, the 95% EVT fit rate is partly built into the method: Algorithm 1 regenerates samples until the CV test passes or times out, and failed cases are excluded as NV. That makes \"EVT fits 95% of cases\" less informative. Third, the abstract says \"5% of cases\" while the introduction and body say \"35%\"—an inconsistency that should have been caught. Fourth, the paper says it uses the threshold approach but fits a GEV; threshold exceedances should be modeled with a generalized Pareto distribution, and the fitting procedure is ambiguous. The 0.05 ECD cutoff is arbitrary, and no significance test on the difference µ_u − µ_p is actually reported.\n\nThe central argument does not hold up as stated, so I would not cite the ECD results in their current form. But the idea is worth engaging with, and the empirical study is extensive enough to deserve referee time.\n\nMy recommendation: do not desk-reject. Send to peer review with a request for major revision. The authors need to redefine ECD on a common tail quantile, reconcile the numbers, clarify the fitting, and report the NV cases. If they do, the paper could be a solid contribution. For a reading group, it would be a good case study in how EVT can look rigorous while hiding a comparability problem.","headline":"The ECD metric is not comparable across groups as defined, so the paper's central empirical claims rest on a threshold artifact, though the underlying idea is worth a serious revision.","tokens_in":28072,"tokens_out":3617,"would_cite":false,"duration_ms":37264,"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":"The paper claims that worst-case AI discrimination can be measured by fitting extreme value distributions to counterfactual outcomes, and that doing so reveals blind spots in average-based fairness evaluation and mitigation.","keywords":["extreme value theory","counterfactual discrimination","fairness testing","worst-case fairness","bias mitigation","generative AI","statistical testing","return levels"],"falsifier":"Recompute ECD for the paper's 40 model-task scenarios using a common threshold rule and sample-size-corrected location estimates, and check whether the reported finding that average-based mitigators increase worst-case discrimination in 35% of cases still holds; if the effect reverses under this re-analysis, the central empirical claim is falsified.","tokens_in":27035,"feed_emoji":"⚖️","tokens_out":7359,"duration_ms":68982,"temperature":0.7,"pith_summary":"Average-case fairness metrics can create a false sense of safety, because discrimination that affects only the extreme tail of outcomes stays invisible when the mean looks fair. The paper proposes extreme counterfactual discrimination (ECD), defined as the difference $\\mu_u - \\mu_p$ between the location parameters of generalized extreme value (GEV) distributions fitted to the upper tail of each group's counterfactual discrimination scores, and claims it quantifies the worst-case disadvantage attributable to group membership. Across four classifiers and nine datasets, the paper reports that the GEV fit is statistically valid in 95% of scenarios, that ECD finds significant tail discrimination in 48% of cases while average metrics find only 25%, and that average-based bias mitigators increase worst-case discrimination in 35% of cases. It also claims that an ECD-guided mitigation strategy improves tail fairness in 90% of cases without degrading average fairness, which would make tail awareness a practical complement to standard fairness tooling.","feed_headline":"Extreme-value metric finds AI bias that averages hide","feed_subtitle":"New EVT-based test shows common bias mitigators worsen worst-case discrimination in 35% of cases.","key_machinery":"The load-bearing machinery is the generalized extreme value (GEV) distribution fitted to the upper tail of counterfactual discrimination scores via the threshold approach, where the location parameter $\\mu$ of each group's fitted GEV is interpreted as the expected extreme discrimination for that group and $\\mu_u - \\mu_p$ defines ECD. To gather enough extreme samples, the paper runs a randomized test-case generation loop that measures counterfactual discrimination, applies an exponentiality test based on the coefficient of variation to a growing number of top-$k$ samples to certify that the tail is exponential or light-tailed, and augments under-represented groups with synthetic samples drawn from a learned generative model. The fitted GEV then supplies return levels, i.e., the expected worst-case discrimination over $m$ future interactions, which is how the paper translates the tail fit into an actionable fairness guarantee.","core_discovery":"The paper's central claim is that the worst-case counterfactual discrimination of a model can be estimated and compared across protected groups by fitting a generalized extreme value (GEV) distribution to the upper tail of the counterfactual discrimination values for each group and taking the difference of the location parameters, $\\mathrm{ECD} = \\mu_u - \\mu_p$. When the fitted shape parameter is zero or negative, the paper treats the fit as valid for extrapolation, yielding return levels that state the expected maximum discrimination within $m$ future interactions with the model. Empirically, the claim is that this procedure works in 95% of the 80 (model, dataset, group) scenarios, that tail discrimination appears in 48% of the 40 task-scenarios whereas average discrimination appears in 25%, and that ECD catches tail discrimination that the conditional value at risk measure misses in 18 of 40 experiments. The paper further claims that average-based bias mitigators increase worst-case discrimination in 35% of cases, that the tail-aware MiniMax-Fairness method increases it in 30% of cases, and that a new ECD-optimizing in-process mitigator improves tail fairness in 90% of its 20 DNN and logistic regression configurations while preserving average fairness.","pith_inferences":["One natural extension the paper leaves implicit is to check whether ECD rankings are stable when a single shared threshold and tail size rule is applied to both groups, since the paper fits each group at its own threshold and with its own sample count.","The same extreme-value machinery could be adapted to continuous protected attributes and regression outcomes, where counterfactual differences are not binary flips and tail events must be redefined.","A direct validation experiment would compare ECD's extrapolated return levels against the maximum counterfactual discrimination observed in a large held-out sample, which the paper itself notes as a lower-bound ground truth."],"forward_implications":["Fairness evaluation should report a tail metric such as ECD alongside average metrics; the paper's data suggest that average-only evaluation misses substantial tail discrimination in about a third of scenarios.","Bias mitigators that optimize average fairness cannot be assumed to help worst-case outcomes; the paper reports that average-based mitigators increase ECD in 35% of its scenarios.","A mitigation objective that directly minimizes ECD, with accuracy as a constraint, can reduce tail discrimination without worsening average fairness, according to the paper's results across 20 DNN and logistic regression configurations.","EVT-based return levels turn worst-case fairness into a concrete operational statement, such as the maximum expected discrimination within the next $m$ queries to a deployed model."],"supporting_citations":[{"why":"Defines counterfactual discrimination and average causal discrimination, the foundation on which ECD is built and the main comparison metric.","marker":"[14]"},{"why":"Supplies the generalized extreme value distribution, threshold method, and return-level concepts used throughout.","marker":"[13]"},{"why":"Provides the coefficient-of-variation exponentiality test used to certify that enough tail samples exist for a valid GEV fit.","marker":"[17]"},{"why":"Introduces the conditional tabular generative models used to synthesize realistic samples for sparse protected groups.","marker":"[18]"},{"why":"One of the average-based bias mitigators whose tail behavior is evaluated and reported to worsen ECD in 35% of cases.","marker":"[20]"},{"why":"Another average-based mitigator evaluated in the tail; part of the same set of results.","marker":"[21]"},{"why":"The tail-aware minimax mitigation baseline that the paper compares against and finds increases ECD in 30% of cases.","marker":"[24]"},{"why":"Introduces the CVaR fairness risk measure that ECD is shown to outperform in detecting tail discrimination.","marker":"[15]"},{"why":"The hyperparameter-optimization framework that the paper extends into its ECD-guided mitigator.","marker":"[42]"}],"fun_headline_variants":["Extreme-value metric catches AI bias that averages miss","Bias mitigators can worsen worst-case AI discrimination","Tail-focus fairness test reveals backfiring bias fixes","Worst-case discrimination metric outperforms average checks","AI fairness: New test targets tail risk, not just means"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the GEV location parameters $\\mu_u$ and $\\mu_p$ can be compared directly across groups even though each group is fitted at its own threshold and from a very different number of tail samples; if the apparent location shift is an artifact of those fitting choices, the ECD metric and the empirical claims built on it do not follow.","fun_headline_variants_meta":{"raw":{"variants":["Extreme-value metric catches AI bias that averages miss","Bias mitigators can worsen worst-case AI discrimination","Tail-focus fairness test reveals backfiring bias fixes","Worst-case discrimination metric outperforms average checks","AI fairness: New test targets tail risk, not just means"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000405,"raw_usage":{"total_tokens":2181,"prompt_tokens":1092,"completion_tokens":1089,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":708,"completion_tokens_details":{"reasoning_tokens":1011}},"tokens_in":708,"tokens_out":1089,"duration_ms":12089,"temperature":1.0,"reasoning_tokens":1011,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:04:52.922322+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute ECD for the paper's 40 model-task scenarios using a common threshold rule and sample-size-corrected location estimates, and check whether the reported finding that average-based mitigators increase worst-case discrimination in 35% of cases still holds; if the effect reverses under this re-analysis, the central empirical claim is falsified.","supporting_citations":[{"cited_title":"Coles, J","cited_arxiv_id":null,"evidence_quote":"Supplies the generalized extreme value distribution, threshold method, and return-level concepts used throughout."},{"cited_title":"Measurement- based worst-case execution time estimation using the coefficient of variation,","cited_arxiv_id":null,"evidence_quote":"Provides the coefficient-of-variation exponentiality test used to certify that enough tail samples exist for a valid GEV fit."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the conditional tabular generative models used to synthesize realistic samples for sparse protected groups."},{"cited_title":"A reductions approach to fair classification,","cited_arxiv_id":null,"evidence_quote":"One of the average-based bias mitigators whose tail behavior is evaluated and reported to worsen ECD in 35% of cases."},{"cited_title":"Fairness risk measures,","cited_arxiv_id":null,"evidence_quote":"Introduces the CVaR fairness risk measure that ECD is shown to outperform in detecting tail discrimination."}],"review_version":1}