{"id":"2d28800c-8ff6-420f-bd80-e0e3469e31df","arxiv_id":"2502.05349","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Contextual Scenario Generation (CSG) trains a neural mapping from context x to K surrogate scenarios via a distributional MMD loss and a task-based surrogate loss with MMD regularization, enabling fast high-quality 2SP decisions.","lead":"A new framework learns a mapping from contextual side-information to a small set of surrogate scenarios for two-stage stochastic programs, so that expensive optimization can be run quickly online. It reports competitive decisions on newsvendor, capacity expansion, portfolio, and substitution problems, but the claimed generalization guarantees are not actually derived.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MMD regularization does not control the loss-net's approximation error where the task-net operates; absent a lambda-selection rule, the reported best-case results do not establish reliable high-quality decisions.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing point: PCSG requires Eψ to be a good approximation of ℓopt over the task-net's output distribution, and MMD regularization is the only mechanism enforcing this, yet it does so only heuristically. My reading agrees and sharpens it: the experiments do not merely show sensitivity to λ; they show catastrophic, orders-of-magnitude failures at small λ (e.g., CVaR K=1 static λ=80: 1.03e7% median gap) that are consistent with the loss error maximization the authors themselves describe. A reviewer cannot tell from the paper whether such failures would occur at the λ a practitioner would choose, because no selection rule is given. The concrete test of nested validation-based λ selection would settle whether the concern is a practical obstacle or merely a theoretical caveat. The secondary issue of the unsupported 'finite-sample generalization guarantees' in the abstract is real (no theorem appears; Section 5 lists it as future work), but it is less central to the algorithmic contribution than the λ-reliability problem. Both are consistent with the reader's CONDITIONAL verdict, so I recommend no change.","tokens_in":34763,"tokens_out":4341,"duration_ms":47265,"concrete_test":"Run a nested validation study on the CVaR experiment: for each trial and each K, select λ on a validation fold by minimizing the mean true task loss ℓopt over held-out contexts, then evaluate the chosen task-net on a separate test fold. If the selected λ frequently yields median gaps comparable to the best λ in Table 5 (e.g., <15%), the concern is mitigated; if it sometimes selects catastrophic λ (gaps >1000%), the method lacks a reliable tuning procedure and the reported best-case results are not reproducible by a practitioner.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.5 (Static Approach) replaces the true task loss ℓopt with a learned surrogate Eψ and solves min_φ (1/n)Σ Eψ(f_φ(x), ω) + λℓMMD(f_φ(x), ω). The central premise is that Eψ approximates ℓopt on the scenario sets f_φ actually produces, and that the MMD term keeps f_φ(x) close to the training distribution used to fit Eψ. Neither is guaranteed: no bound relates the MMD distance to the approximation error of Eψ, and the regularization strength λ is a free hyperparameter with no data-driven selection rule. The paper itself documents the failure mode ('loss error maximization', Section 3.5), and the experiments confirm it: CVaR Table 5 reports median optimality gaps of 1.03e7% for Static K=1 λ=80 and 2.60e7% for λ=320; the newsvendor results for λ=0.1 are likewise poor (Fig. 6). At larger λ the task-net is effectively pinned to the MMD solution, so the task-based component adds little. The method's success therefore depends on an unpredictable 'sweet spot' for λ, with no procedure given to find it from data. Without this, the headline claim of a broadly applicable CSG framework yielding high-quality decisions is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes contextual scenario generation (CSG) for two-stage stochastic programs: given side information x, learn a mapping f that outputs K surrogate scenarios, which are then used to solve the SAA problem on those K scenarios. Two methods are developed: a distributional approach (DCSG) that minimizes an MMD distance between the empirical distribution of generated scenarios and the conditional distribution of uncertainty, and a task-based approach (PCSG) that trains a task network using a learned neural surrogate Eψ of an optimistic task loss ℓopt, regularized by the MMD term. The optimistic loss handles non-unique solutions of the SAA problem, and Proposition 2 motivates the permutation-invariant architecture of the loss network. Experiments on newsvendor, CEP1 capacity planning, CVaR portfolio optimization, and a multidimensional newsvendor with substitution compare DCSG, static PCSG, dynamic PCSG, expected-value, and quantile-regression benchmarks, reporting optimality gaps over a grid of regularization parameters λ.","tokens_in":35075,"tokens_out":6512,"duration_ms":70856,"significance":"If the central claim is established, this would be a valuable contribution: it offers a way to replace expensive scenario generation and reduction with a cheap neural forward pass, while keeping the downstream 2SP solve tractable. The optimistic relaxation of the bi-level scenario-generation problem is a sensible way to handle non-unique SAA solutions, and the paper covers four structurally different 2SP classes. The work is honest about a known failure mode (loss error maximization) and about the mixed performance of the dynamic approach in the MNV experiment. However, the paper advertises finite-sample generalization guarantees that do not appear anywhere in the text, and the reported success depends on an ex post selection of the regularization strength λ; no data-driven rule is given. These issues directly affect the reliability of the headline claim.","major_comments":[{"comment":"The abstract states “We provide finite-sample generalization guarantees,” but no such theorem, proposition, or bound appears in Sections 3–4, and Section 5 explicitly defers this: “Future work could explore … understanding finite sample learning bounds and the generalization/stability properties of the proposed methodologies.” This is a direct contradiction between the advertised contribution and the content of the paper. The claim should either be substantiated with a precise statement (e.g., a generalization bound for the empirical minimizer of LMMD or of the static PCSG objective) or removed from the abstract.","section":"Abstract and Section 5"},{"comment":"The central premise of PCSG is that the learned loss-net Eψ approximates ℓopt on the distribution of scenario sets actually produced by the task-net, and that the MMD regularizer prevents the task-net from leaving that well-approximated region. Neither is guaranteed: no bound links the MMD distance to the approximation error of Eψ, and λ is a free hyperparameter with no data-driven selection rule. The paper itself documents the failure mode (“loss error maximization”) and the experiments confirm it: CVaR Table 5 reports median optimality gaps of 1.03e7% for Static K=1 with λ=80 and 2.60e7% with λ=320; newsvendor Figure 6 shows poor performance at λ=0.1. At large λ the method collapses toward the MMD solution, so the task-based component adds little. The reported best results are selected over a grid of λ per problem and per K, which does not establish that the method can be deployed without a priori knowledge of a good λ. A practical model-selection procedure for λ (e.g., cross-validation on the task loss) or an analysis that justifies a default or adaptive choice is needed.","section":"Section 3.5, Static Approach and Tables 2, 5, 8"},{"comment":"The optimistic relaxation replaces min_{y∈Y*(f(x))} E_{ω|x}[h(y)+Q(y,ω)] with E_{ω|x}[min_{y∈Y*(f(x))} h(y)+Q(y,ω)], which is a lower bound because it allows the choice of y ∈ Y* to depend on the realized ω. Training on this lower bound does not automatically yield a low true objective; the gap between the two quantities is not quantified, and nothing in the method controls it. The paper notes the inequality but does not investigate how loose it is on the tested instances. Since the evaluation uses the true 2SP objective with out-of-sample conditional distributions, a loose optimistic bound could give the task-net a misleading training signal. I would like to see, at least on the validation instances, a comparison between the optimistic training loss and the actual “best-in-hindsight” task loss (or a discussion of why the optimism is benign for the problem classes considered).","section":"Section 3.4, Eq. (Opt-PCSG')"},{"comment":"Proposition 2 gives an exact sum-decomposable representation of ℓopt under a continuity assumption, but it provides no approximation guarantee for the finite-capacity ReLU networks used in the loss-net, and the continuity assumption is not verified for the mixed-binary problems (CVaR and MNV) in the experiments. The paper acknowledges that “we leverage a continuous approximation to ℓopt, irrespective of whether ℓopt is continuous,” but this is a heuristic step. The role of Proposition 2 should be clarified: it motivates the architecture, but it does not by itself justify the reliability of the loss-net approximation. I recommend stating explicitly that the approximation step is heuristic and unsupported by guarantees, and discussing what could go wrong when the loss-net is inaccurate (beyond the already documented loss error maximization).","section":"Section 3.5, Proposition 2"}],"minor_comments":[{"comment":"There are several typos: “Lipshitz” should be “Lipschitz” (Section 3.1); “PSCR” appears in the CEP1 results text (Section 4.2) where “PCSG” is meant; figure labels in Appendix C contain “I sta ces” and “% f Instances”; “ReLu” should be “ReLU” in several places; the caption of Figure 3 ends with the single word “Embedded” and appears incomplete.","section":"Throughout"},{"comment":"The discussion of the energy kernel and the choice of MMD is clear, but the phrase “parameter-free kE” could be misread: the kernel has no free parameters, yet the method still requires choosing K and the neural architecture. Please clarify that “parameter-free” refers only to the kernel.","section":"Section 3.2 and Appendix A"},{"comment":"The observation that the dynamic approach failed in the MNV experiment because the loss-net architecture did not generalize is valuable, but it is placed only in the results discussion. Consider moving this to a limitations paragraph in Section 5, since it directly informs the applicability of the dynamic method.","section":"Section 4.4, last paragraph"},{"comment":"In the proof, after applying Proposition 1 from Tabaghi and Wang, the function ρ is defined by ρ(ζ̂, ω̂) = ρ̃(Kζ̂, ω̂). This definition is only needed for multisets of size exactly K, and the text could state more clearly that the extension to smaller multisets is irrelevant for the application at hand.","section":"Appendix B, proof of Proposition 2"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid formulation and an extensive experimental study, but the mismatch between the abstract's claim of finite-sample guarantees and the actual content is a serious issue that the authors must resolve. The lambda-sensitivity problem is also central: the reported results are best-over-grid, and without a principled lambda selection the method does not yet deliver the advertised reliability. I would not rule out acceptance after a major revision, but the authors need to either provide the promised guarantees, remove the claim and present the method as a heuristic with clearly stated limitations, and add a practical selection rule or a sensitivity analysis that shows performance is robust across a reasonable range of lambda."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper deserves a serious referee, but only after the authors remove or substantiate the finite-sample guarantee in the abstract and address the lambda-sensitivity that their own experiments document. The core idea is genuinely new: a learned mapping from context to a small scenario set for 2SPs, with a distributional MMD branch and a task-based branch using an optimistic loss computed by solving an auxiliary 'Opt-Search' problem. That is not just repackaged predict-then-optimize; generating a set of K scenarios and evaluating via (ζ-SAA) is a different object than point prediction, and the optimistic treatment of non-unique solutions is sensible. The MMD derivation is clean, Proposition 2 gives a representation theorem from Tabaghi-Wang, and the dynamic training algorithm is honestly credited to Zharmagambetov et al. The four experiments span LP, MIP, and mixed-binary structures, and the timing tables make the offline-training tradeoff concrete. Credit where due: the paper reports its failures (loss error maximization, the MNV dynamic collapse) rather than hiding them.\n\nThe soft spots are exactly where the reader puts them. First, the abstract promises 'finite-sample generalization guarantees' and no theorem appears; the conclusion defers them to future work. That is a straightforward overclaim and must be fixed. Second, the static approach relies on Eψ approximating ℓopt on the task-net's outputs, and MMD regularization is the only leash. The stress-test is correct: no bound connects MMD distance to loss-net error, and λ is a free parameter with no selection rule. Tables 5 and 6 show median gaps of 10^7% at small λ; at large λ the method collapses to MMD. So the reported best cases are real but not yet reproducible by a practitioner without per-problem hand-tuning. Third, the claim of 'broadly applicable' is undercut by the MNV dynamic failure, though they honestly report it.\n\nThe math and data look solid as far as they go. I would not cite this version, but a revised version addressing the abstract and λ-selection would be citable. Bring to reading group? Maybe — the problem formulation is worth discussing. Recommend: send to peer review, major revision.","headline":"Solid, honest first framework for contextual scenario generation, but the abstract overpromises finite-sample guarantees and the missing lambda-selection rule means the headline results are not yet reproducible.","tokens_in":35558,"tokens_out":2060,"would_cite":false,"duration_ms":22714,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C15","90C11","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes contextual scenario generation: learn a mapping from context to K surrogate scenarios so that solving the two-stage stochastic program on those scenarios yields near-optimal decisions at low computational cost.","keywords":["contextual scenario generation","two-stage stochastic programming","maximum mean discrepancy","task-based learning","surrogate scenarios","end-to-end optimization","contextual optimization"],"falsifier":"Run the task-based method with lambda = 0 on a held-out context distribution and measure the MMD distance between the produced scenario sets and the loss-net training scenarios; the paper's own newsvendor and CVaR tables show that small lambda produces enormous optimality gaps, so if the central claim is meant to hold for any lambda, this is a direct refutation. A systematic sweep of lambda across K would map the regime where the method degenerates and reveal whether any fixed regularization reliably prevents loss error maximization across problem classes.","tokens_in":34567,"feed_emoji":"🎯","tokens_out":5692,"duration_ms":59960,"temperature":0.7,"pith_summary":"The paper proposes contextual scenario generation (CSG): instead of estimating the conditional distribution of uncertainty and then sampling and reducing a large scenario set at decision time, learn a neural mapping from observed context to a small, user-specified number K of surrogate scenarios. Solving the two-stage stochastic program on just those K scenarios then gives first-stage decisions whose true conditional expected cost is close to optimal. The paper develops two complementary training losses: a distributional loss based on kernel maximum mean discrepancy with the energy distance, and a task-based loss that differentiates through a learned surrogate of the optimistic downstream objective, regularized by the same MMD term to keep the task network from exploiting errors in the surrogate. Experiments on newsvendor, capacity planning, CVaR portfolio selection, and a multidimensional newsvendor with substitution show that task-based CSG with suitable regularization can outperform the distributional baseline and the expected-value solution, and can unlock the value of the stochastic solution from very few scenarios.","feed_headline":"Context learns K scenarios; two-stage decisions stay near-optimal","feed_subtitle":"A neural forward pass replaces sampling and scenario reduction, cutting decision-time cost while keeping solution quality.","key_machinery":"The key machinery is the loss-net E_psi, defined by mean aggregation over encoded surrogate scenarios: E_psi(zeta_1...K, omega) = Psi_2((1/K) sum_k Psi_1(zeta_k), Psi_1(omega)), with Psi_1 and Psi_2 feedforward ReLU networks. Its purpose is to approximate the optimistic task-based loss l_opt(zeta_1...K, omega), which is computed by solving the K-scenario SAA and then an additional recourse search over its optimal solution set. Proposition 2, built on a multiset sum-decomposition result, shows that l_opt has exactly this functional form with continuous embedding and combining maps, so the architecture is a universal approximator for the task loss. The task-mapping f_phi is a neural network trained to minimize the loss-net output plus an MMD regularization term with the energy-distance kernel; the dynamic variant iteratively refines the loss-net on replay-buffer samples generated around the current f_phi.","core_discovery":"The central claim is that there exists an efficiently evaluated mapping f : X to $\\Omega$^K such that solving the sample-average approximation on f(x) is a high-quality proxy for solving the true contextual two-stage problem under the conditional distribution P_{omega|x}. The paper establishes this in two ways: DCSG minimizes the expected squared MMD between the empirical distribution on f(x) and P_{omega|x}, using the parameter-free energy kernel; PCSG minimizes the optimistic task loss l_opt(zeta, omega) = min_{y in Y*(zeta)} h(y) + Q(y, omega), approximated by a permutation-invariant neural network to avoid the sparse and uninformative gradients of the true recourse function. The optimistic formulation handles non-unique optimal solutions to the SAA problem, and the MMD regularization term $\\lambda$*l_MMD prevents the failure mode the paper calls loss error maximization, in which the task network drives scenarios outside the region where the surrogate loss is accurate. Empirically, with a suitable choice of $\\lambda$, the task-based approaches beat the distributional baseline for small K and achieve median optimality gaps below 10 percent in the CVaR problem with K = 40, while reducing decision-time computation to a forward pass plus a K-scenario solve.","pith_inferences":["An extension the paper leaves implicit is that the same loss-net plus MMD-regularized task-net recipe could be applied to multistage stochastic programs, replacing the two-stage optimistic loss with stage-wise recourse evaluations.","Because the energy-distance MMD with K = 1 reduces to least-squares conditional mean estimation, the distributional approach is a natural initialization for the task-based approach; a testable implication is that MMD-initialized task networks should dominate random initialization when lambda is moderate.","The abstract advertises finite-sample generalization guarantees, but the full text as posted defers such bounds to future work; a reader should verify whether a later version states a concrete sample-complexity theorem before relying on that claim.","A natural ablation is to replace the energy kernel with Sinkhorn divergence or a conditional Wasserstein distance (both discussed in the appendix); if the performance gap between task-based and distributional approaches persists, the bottleneck is the surrogate loss rather than the distributional metric."],"forward_implications":["At decision time CSG replaces conditional-density estimation, sampling, and scenario reduction with one forward pass through the task network, so per-context cost is independent of the original sample size M.","For small K (K <= 5), task-based methods with suitable MMD regularization can outperform both the distributional baseline and the expected-value solution, meaning a stochastic solution is obtainable from very few scenarios.","Because the optimistic loss is computed by solving a K-scenario SAA and an extra recourse problem, the method applies to two-stage LPs, MIPs with nonlinear or integer elements, and problems without unique optimal solutions.","The dynamic refinement variant improves on the static variant when the loss-net keeps generalizing on new scenario sets, but it requires repeated evaluation of the optimistic loss and can degrade when the loss-net fails to generalize.","Training costs are paid offline; the timing analysis shows that hundreds of CVaR instances can be solved in about 33 seconds with K = 40 while keeping a median optimality gap around 5 percent."],"supporting_citations":[{"why":"Provides the conditional-MMD metric and Theorem 1, establishing that the distributional objective is zero if and only if the predicted conditional distribution matches the true one.","marker":"[27]"},{"why":"Supplies the universal multiset sum-decomposition used in Proposition 2 to show the optimistic task loss is representable by the loss-net architecture.","marker":"[35]"},{"why":"Introduces the surrogate-loss training paradigm and the phenomenon of loss error maximization that motivates the MMD regularization in PCSG.","marker":"[11]"},{"why":"Provides the permutation-invariant neural architecture for two-stage stochastic programs that the loss-net design builds on.","marker":"[3]"},{"why":"Gives the kernel mean embedding and squared MMD formula that underpin the distributional loss and its sample estimate.","marker":"[25]"},{"why":"Establishes the characteristic property of the energy distance, justifying the parameter-free kernel used for the MMD regularization.","marker":"[30]"},{"why":"Supplies the interchangeability result used to convert the pointwise bi-level scenario generation objective into the population-level Opt-PCSG minimization.","marker":"[33]"}],"fun_headline_variants":["Context maps to K scenarios, keeping two-stage decisions near-optimal","Neural scenario generation from context speeds up two-stage optimization","Learn a scenario set from context for near-optimal 2SP decisions","From context to a small scenario set: fast and near-optimal 2SP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The task-based pipeline depends on the loss-net being accurate on the distribution of scenario sets that the task-net actually produces, and on the MMD penalty being strong enough to keep the task-net inside that well-approximated region; when the penalty is too small, the task-net exploits surrogate-loss errors and optimality gaps explode.","fun_headline_variants_meta":{"raw":{"variants":["Context maps to K scenarios, keeping two-stage decisions near-optimal","Neural scenario generation from context speeds up two-stage optimization","Learn a scenario set from context for near-optimal 2SP decisions","From context to a small scenario set: fast and near-optimal 2SP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000357,"raw_usage":{"total_tokens":1963,"prompt_tokens":997,"completion_tokens":966,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":889}},"tokens_in":613,"tokens_out":966,"duration_ms":9730,"temperature":1.0,"reasoning_tokens":889,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T19:41:01.071743+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the task-based method with lambda = 0 on a held-out context distribution and measure the MMD distance between the produced scenario sets and the loss-net training scenarios; the paper's own newsvendor and CVaR tables show that small lambda produces enormous optimality gaps, so if the central claim is meant to hold for any lambda, this is a direct refutation. A systematic sweep of lambda across K would map the regime where the method degenerates and reveal whether any fixed regularization reliably prevents loss error maximization across problem classes.","supporting_citations":[{"cited_title":"In: International Conference on Algorithmic Learning Theory, pp","cited_arxiv_id":null,"evidence_quote":"Supplies the universal multiset sum-decomposition used in Proposition 2 to show the optimistic task loss is representable by the loss-net architecture."},{"cited_title":"Advances in Neural Information Proce ssing Systems 36 (2024)","cited_arxiv_id":null,"evidence_quote":"Introduces the surrogate-loss training paradigm and the phenomenon of loss error maximization that motivates the MMD regularization in PCSG."},{"cited_title":"Advances in Neural Information Proces sing Systems 35, 23992–24005 (2022) 42","cited_arxiv_id":null,"evidence_quote":"Provides the permutation-invariant neural architecture for two-stage stochastic programs that the loss-net design builds on."},{"cited_title":"Journal of Machine Learning Research 13(25), 723–773 (2012)","cited_arxiv_id":null,"evidence_quote":"Gives the kernel mean embedding and squared MMD formula that underpin the distributional loss and its sample estimate."},{"cited_title":"Statistics & Probability Letters 82(12), 2278–2282 (2012) https://doi.org/10.1016/j.spl.2012","cited_arxiv_id":null,"evidence_quote":"Establishes the characteristic property of the energy distance, justifying the parameter-free kernel used for the MMD regularization."}],"review_version":1}