{"id":"1985553f-f8c8-45f0-8a56-a6f4fadb88fe","arxiv_id":"2509.08942","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A min-max-sup extension of Group DRO that adds a Wasserstein ball around each group's empirical distribution, with a descent-mirror-ascent algorithm and Adult income experiments.","lead":"The paper proposes combining Group DRO with per-group Wasserstein distributional robustness, and reports better worst-group accuracy on the Adult income dataset. The convergence theorem is not actually proven, and the reported gains come from hyperparameters selected on the test data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem A.1 does not apply to the implemented Algorithm 2: the inner supremum is only approximated by T_rob unprojected gradient steps, and the experimental γ=10^-4 violates the γ>L_zz condition required for a finite L_f.","rationale":"The paper's own contribution is a 'provably convergent' algorithm for the nested group-DRO problem. That proof is the load-bearing element; if it does not cover the implemented algorithm, the central claim fails regardless of the empirical table. The reader's weakest assumption is exactly the one identified here: exact inner supremum and γ>L_zz are assumed by the theory but not satisfied or verified by Algorithm 2. The L_zz check above would make this concrete: a single computation showing L_zz > 1e-4 would demonstrate that the theorem's smoothness hypothesis is violated. The empirical concerns (single dataset, γ selected on the test set, no code) are additional, but they are not needed to justify the verdict. Since the reader already reached REJECT and this stress-test reinforces that conclusion, no adjustment is needed.","tokens_in":16308,"tokens_out":15282,"duration_ms":165241,"concrete_test":"Implement the exact two-layer ELU MLP from Section 4, train it on the standardized Adult data under the paper's setup, and estimate L_zz empirically as max over at least 10^4 random pairs (z,z') from the training set of ||∇_z L(θ;z) − ∇_z L(θ;z')|| / ||z − z'||. If this maximum exceeds γ=10^-4, then the hypothesis γ>L_zz needed by Lemma A.5 is false for the reported experiments, so Theorem A.1 cannot be invoked.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central convergence claim (Theorem A.1) is for P(θ)=max_q Σ q_g F_g(θ), where F_g is the exact expectation of the inner supremum over z. Algorithm 2, lines 4–6, replaces that supremum with z_i from Algorithm 1 after T_rob=100 steps of unprojected gradient ascent, and no bound on the approximation error (or on finite-sample averaging in line 8) is given. Theorem A.1 also inherits Lemma A.5's requirement γ > L_zz, because L_f = L_θθ + L_θz L_zθ/(γ−L_zz) is finite only then. For the two-layer ELU network used in Section 4, L_zz is never estimated, and the reported experiments fix γ=10^-4, which is orders of magnitude below any plausible Lipschitz constant of ∇_z L for that architecture on standardized features; at best the assumption is unverified and at worst L_f is undefined. The proof is also not supplied: Lemma A.8 imports a GDA result from [Lin et al. 2020], and Theorem A.1 cites an unnamed theorem in [Sinha et al. 2017] rather than deriving the mirror-ascent case. Thus the paper's assertion that Algorithm 2 provably solves (5) is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a group DRO framework that adds per-group Wasserstein distributional uncertainty to the standard Group DRO objective. The formulation is the nested problem (5): min over model parameters, max over group weights, and, within each group, a supremum over distributions in a Wasserstein ball around the empirical group distribution. The authors propose Algorithm 2, which alternates gradient ascent on adversarial perturbations (Algorithm 1), mirror ascent on group weights, and gradient descent on the model parameters. They claim a convergence guarantee (Theorem A.1) for the corresponding function P(θ) = max_q Σ_g q_g F_g(θ), where F_g is the exact expected robust group loss. Experiments on the Adult dataset, including a constructed education marginal shift and multiple test environments, report that the method improves worst-group accuracy and reduces accuracy range compared with ERM, DRO, and Group DRO.","tokens_in":16576,"tokens_out":3260,"duration_ms":37945,"significance":"If the convergence guarantee and empirical results were fully supported, the paper would address a genuine gap: existing Group DRO methods typically assume exact knowledge of each group's distribution, while the proposed nested formulation explicitly models within-group distributional uncertainty. The problem setup is well motivated, and the multi-environment Adult evaluation is a sensible stress test. However, the central theoretical claim is not established: Theorem A.1 relies on an unnamed theorem in a prior paper, and the implemented algorithm does not match the theorem's assumptions because the inner supremum is approximated by unprojected gradient ascent without error control. The experimental validation is also weakened by selection of the robustness parameter on the test data and by the absence of a second promised dataset. The contribution is therefore significant in conception but not currently supported in execution.","major_comments":[{"comment":"The convergence proof is not supplied. The proof of Theorem A.1 states that it follows 'the same steps to combine lemmas A.8, A.9, and A.10 as in Theorem ... in [Sinha et al., 2017]', but no theorem number is given. Lemma A.8 similarly imports a GDA result from [Lin et al., 2020] without a derivation. Since the convergence guarantee is the paper's main theoretical contribution, this is load-bearing and cannot be checked. The authors need to provide a complete proof or a precise, stated theorem from the cited work together with a verification of the conditions.","section":"Appendix A.3, Theorem A.1"},{"comment":"Theorem A.1 concerns the exact function F_g(θ) = E sup_z ϕ(f_θ; (x,y), z), i.e., the expected value of the exact supremum. Algorithm 2 instead sets z_i = argmax via Algorithm 1, which performs only T_rob=100 unprojected gradient ascent steps (Section 4.1 reports T_rob=100). No bound is provided for the approximation error between the T_rob-step iterate and the true supremum, nor for its effect on the gradient or on the Moreau-envelope stationarity measure. Consequently the stated convergence result does not apply to the implemented algorithm. The paper must either analyze the inexact inner problem or modify the algorithm so that the exact supremum is used or provably approximated.","section":"Algorithm 2, lines 4-6; Theorem A.1"},{"comment":"The smoothness constant L_f = L_θθ + L_θz L_zθ / [γ - L_zz]_+ is finite only if γ > L_zz. Lemma A.6 also assumes existence of the argmax for γ > L_zz. The authors never estimate L_zz for the two-layer ELU network used in the experiments, and the reported experiments use γ = 10^-4 (Table 1), with a sweep starting at 10^-4. For a neural network on standardized features, the Lipschitz constant of the loss gradient with respect to z is not plausibly below 10^-4. Thus the assumptions of the convergence theorem are at best unverified and very likely violated in the configuration whose results are reported.","section":"Lemma A.5 and Section 4.1"},{"comment":"Section 4 states that the method is evaluated on 'real-world datasets' (plural), but only the Adult dataset is used; Section 4.2 is a set of constructed test environments on the same Adult data, not a second dataset. More importantly, γ is swept over a wide range and the 'best' value is selected (Section 4.1: 'select the best value via fine-tuning'; Figure 5: 'best-performing γ'), and the final tables report numbers at that selected value. If the selection is made using the test sets that are also used for evaluation, the comparison is fitted to the evaluation data. The authors need a validation-set selection procedure and should report results for a fixed, pre-specified γ or over the whole sweep without test-based selection.","section":"Section 4, Table 1, Figure 5"}],"minor_comments":[{"comment":"The sentence 'we sweep γ ... to select the best value via fine-tuning' is ambiguous. Please specify which data split is used for tuning and how many seeds are used for selection.","section":"Section 4.1"},{"comment":"Line 8 uses the notation 1/N_g and appears to average over the full group dataset. If this is the full-batch gradient, state it explicitly; if minibatches are used, the finite-sample error must be addressed.","section":"Algorithm 2, line 8"},{"comment":"The column header 'Ours (γ=10^-4)' is acceptable, but the header 'Group' is an abbreviation for Group DRO and could be confused with the group index. Please rename for clarity.","section":"Table 1"},{"comment":"The text says 'real-world datasets from the field of finance', but the Adult dataset is not a finance dataset. Please correct the description.","section":"Section 4 opening"},{"comment":"There are several typographical issues, e.g., 'Assimptions' in Lemma A.10 and 'envolope' in the convergence appendix. Typesetting of the many cross-references to equations could also be improved.","section":"Various"}],"recommendation":"reject","confidential_remarks":"The proposed formulation is interesting and potentially relevant to the journal, but the central convergence theorem is not proven, the implemented algorithm is not covered by the stated assumptions, and the experimental protocol allows test-set-based hyperparameter selection. These are load-bearing issues that would require substantial additional theoretical and experimental work to resolve; in the current state I cannot recommend publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: the paper puts a known pair of tools together in a sensible way and gets decent numbers on Adult, but the convergence proof doesn't cover the algorithm that was actually run, and the experimental comparison is weakened by test-set-level tuning. Neither issue kills the idea, but both need to be fixed before I'd trust the paper.\n\nWhat's genuinely there: the formulation (5) is a natural composition of Group DRO with per-group Wasserstein ambiguity sets. The authors correctly note that existing methods handle either across-group reweighting (Sagawa et al.) or within-group shift (Sinha et al.) but not both. The mirror-ascent over group weights and the use of the Lagrangian surrogate for the inner supremum are reasonable starting points. The Adult experiments show a consistent improvement over ERM, DRO, and Group DRO, and the multi-environment extension is a nice way to stress-test behavior under covariate shift.\n\nWhere it falls down: Theorem A.1 claims an epsilon-stationary point for P(theta) = max_q sum_g q_g F_g(theta), where F_g is the exact supremum in Proposition 3.1. Algorithm 2 replaces that supremum with T_rob=100 unprojected gradient ascent steps (Algorithm 1) and then averages over a finite batch. There is no bound on the approximation error from either source. The proof itself is not self-contained: Lemma A.8 imports a GDA result from [Lin et al., 2020] and Theorem A.1 references 'Theorem ...' in [Sinha et al., 2017] without naming it. So the central guarantee is unsupported for the implemented method. Second, Lemma A.5 requires gamma > L_zz for L_f to be finite; L_zz is never estimated for the two-layer ELU net, and the experiments fix gamma=1e-4, which is far below any plausible Lipschitz constant for nabla_z L on standardized features. At best the assumption is unverified; at worst L_f is undefined and the theorem is vacuous. Third, the empirical section promises real-world datasets but only Adult appears, and the gamma sweep is done on the test set to pick the best result. That makes the reported improvements partly self-selected. There is also no code.\n\nThese are load-bearing flaws, but they are fixable. The combination is worth studying; the current version just doesn't deliver the guarantees it claims. I'd send it to peer review in the hope that the authors can close the approximation-theory gap and run a cleaner empirical study, but as it stands the paper's central claims are not credible.\n\nFor myself: I wouldn't cite this in its current form, but I'd be interested in a revised version that actually proves the bound or changes the algorithm to match the theorem.","headline":"Reasonable composition of Group DRO and per-group Wasserstein robustness, but the convergence proof doesn't cover the implemented algorithm and the test-set-level tuning of gamma undercuts the empirical claims.","tokens_in":17064,"tokens_out":4694,"would_cite":false,"duration_ms":44224,"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":"This paper argues that classical Group DRO ignores uncertainty in each group's own data distribution, and shows that adding per-group Wasserstein balls yields a tractable min-max-sup problem with better worst-group accuracy.","keywords":["group distributionally robust optimization","Wasserstein distance","distribution shift","worst-group accuracy","min-max optimization","fairness","adversarial training"],"falsifier":"Estimate L_zz empirically for the two-layer ELU network on the Adult features and check whether the deployed gamma values (including 10^-4) satisfy gamma > L_zz; alternatively, increase T_rob until the inner penalized loss stabilizes and see whether the reported worst-group accuracy and spread change materially. Either test would show whether the convergence theorem covers the method as implemented.","tokens_in":16192,"feed_emoji":"⚖️","tokens_out":5279,"duration_ms":53976,"temperature":0.7,"pith_summary":"Machine learning models often degrade on underrepresented groups when training data are heterogeneous and group distributions are not known exactly. Classical Group DRO protects against the worst mixture of groups but assumes each group's training distribution is accurate. This paper proposes to embed each group's empirical distribution in a Wasserstein ball, forming a nested min-max-sup objective that is robust to both group reweighting and within-group shift. The paper provides a convergence guarantee for a gradient descent-mirror ascent algorithm with an inner gradient ascent step for the adversarial perturbation, and reports on the Adult income dataset that the method raises worst-group accuracy from 0.56 (Group DRO) to 0.61 while cutting the accuracy range from 0.26 to 0.19.","feed_headline":"Worst-group accuracy rises from 0.08 to 0.61 with group-level DRO","feed_subtitle":"Nesting Wasserstein ambiguity sets inside each group's loss fixes a blind spot in classic Group DRO.","key_machinery":"The central object is the nested objective (5): min over parameters theta of max over group weights q of the sum over groups of q_g times the per-group robust loss, where each robust loss is a supremum over distributions within a 1-Wasserstein ball of radius eps_g around the group's empirical distribution. The enabling identity is Proposition 3.1, a Lagrangian relaxation that converts the inner supremum into an expectation over training samples of a penalized loss L(f_theta; z') - gamma c(z, z'), which can be maximized by gradient ascent. Around this sit a mirror-ascent update for the group weights, a gradient descent update for the parameters, and a Moreau-envelope stationarity argument tha","core_discovery":"The paper's central discovery is that the two forms of uncertainty—uncertainty in the mixture of environments and uncertainty in each environment's own distribution—can be handled in a single optimization problem. The proposed objective minimizes over model parameters the worst-case weighted sum of per-group robust losses, where each group's robust loss is the supremum of the expected loss over a Wasserstein ball around the group's empirical distribution. The authors show that this inner supremum can be computed via a Lagrangian relaxation that turns it into an expectation of a penalized loss, making it tractable for nonconvex models. The paper proves that the resulting three-step gradient a","pith_inferences":["Because gamma=0 recovers standard DRO and gamma tending to infinity recovers Group DRO, the framework implicitly defines a continuum of robustness levels; the paper leaves unexplored how to choose gamma per group according to estimated shift severity.","The convergence theorem assumes the inner supremum is computed exactly, but the algorithm uses a fixed number of gradient ascent steps with no projection; an empirical check of the ratio gamma/L_zz on the actual network would show whether the reported results operate in the theorem's regime.","A natural extension is to test the method under multiple simultaneous covariate shifts (age, occupation, region) rather than a single education shift, which would clarify whether the observed advantage is specific to the constructed split.","The reported accuracy gains are relative to methods that ignore one of the two uncertainty sources; a more direct comparison would be to a baseline that uses the same inner adversarial training but with a fixed uniform q, isolating the contribution of the mirror ascent over group weights."],"forward_implications":["Models trained with this objective are protected against two failure modes at once: shift in the group mixture and drift inside each group's data distribution, both without requiring knowledge of the true group distributions.","The Lagrangian relaxation makes the inner worst-case distribution search tractable for nonconvex models such as neural networks, where dual formulations of the robust loss are not available.","The method stays stable across gamma values from 10^-4 to 10^1 in the Adult experiments, so the penalty parameter does not require delicate tuning.","On the Adult income task with an education covariate shift, the method achieves the highest average accuracy (0.715), highest worst-group accuracy (0.613), and lowest accuracy range (0.193) among the four compared methods.","Across constructed test environments ranging from 90-10 to 10-90 education splits, the method maintains worst-group accuracy around 0.60-0.63, roughly half the disparity of Group DRO."],"fun_headline_variants":["Per-group Wasserstein balls lift worst-group accuracy to 0.61","Nested Wasserstein ambiguity sets fix Group DRO blind spot","A single DRO for both group mixture and within-group noise","Worst-group DRO with Wasserstein balls hits 0.61 accuracy"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proof of convergence requires that each step's inner worst-case distribution be found exactly (or with negligible error), but the implemented algorithm approximates it with a finite, unprojected gradient ascent and never verifies that the penalty parameter gamma exceeds the loss's input-space smoothness constant L_zz.","fun_headline_variants_meta":{"raw":{"variants":["Per-group Wasserstein balls lift worst-group accuracy to 0.61","Nested Wasserstein ambiguity sets fix Group DRO blind spot","A single DRO for both group mixture and within-group noise","Worst-group DRO with Wasserstein balls hits 0.61 accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000982,"raw_usage":{"total_tokens":3973,"prompt_tokens":681,"completion_tokens":3292,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":425,"completion_tokens_details":{"reasoning_tokens":3225}},"tokens_in":425,"tokens_out":3292,"duration_ms":28585,"temperature":1.0,"reasoning_tokens":3225,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T19:56:16.774899+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Estimate L_zz empirically for the two-layer ELU network on the Adult features and check whether the deployed gamma values (including 10^-4) satisfy gamma > L_zz; alternatively, increase T_rob until the inner penalized loss stabilizes and see whether the reported worst-group accuracy and spread change materially. Either test would show whether the convergence theorem covers the method as implemented.","supporting_citations":[],"review_version":1}