{"id":"6bd2330f-a135-4451-ba86-4b0006e6712b","arxiv_id":"2607.05830","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":6,"one_line_summary":"A unified decision-focused framework trains VAE, GAN, and diffusion models to generate correlated scenarios for distributionally robust grid dispatch, reducing operational cost by 0.80–2.02% over accuracy-oriented methods.","lead":"This paper trains generative models (VAE, GAN, diffusion) to produce power-grid uncertainty scenarios that minimize downstream dispatch cost rather than forecasting error, reducing operational cost by 0.80–2.02%. A smart generalist might read it to understand how decision-focused learning can bridge the gap between machine-learning predictions and the operational decisions they serve.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The 0.80%–2.02% improvement claim is under-supported: several settings show improvements below 0.50%, no significance testing is provided, and the abstract appears to selectively report only the better results.","rationale":"The reader correctly identified the KKT nonsingularity assumption as an unverified mathematical premise, and this is a legitimate concern. However, I view it as less load-bearing for the central empirical claim than the absence of statistical validation. The OptNet framework [28] is well-established, and while LP degeneracy can cause Jacobian singularity, this is a known issue with standard mitigations (regularization, perturbation). In contrast, the experimental evidence directly underpins the headline claim: without significance testing, we cannot distinguish real improvement from noise, especially for the settings showing 0.11%–0.49% improvements. The selective reporting in the abstract (claiming 0.80%–2.02% when the full range is 0.11%–2.09%) further weakens the claim's precision. The CONDITIONAL verdict is appropriate: the framework is well-motivated and internally consistent, the code is available, and the joint-forecasting results are directionally promising, but the empirical claim needs stronger validation before it can be fully accepted. The reader's additional concerns about unreported hyperparameters (n_tru, λ, ν, τ_g) and the single test system are also valid and compound the experimental validation gap.","tokens_in":20566,"tokens_out":4979,"duration_ms":309821,"concrete_test":"Re-run all eight forecasting settings in Table I with at least 5–10 random seeds, reporting mean ± standard deviation for each cell. Perform paired statistical tests (e.g., Wilcoxon signed-rank) for DF(Selector) vs. AO in each column. If the improvements for GAN-Separate (0.49%) or Diff-Separate (0.31%) are not statistically significant at p<0.05, the abstract's '0.80%–2.02%' claim should be revised to reflect that the benefit is concentrated in joint-forecasting settings only. Additionally, test on at least one larger system (e.g., IEEE 118-bus) to assess generalizability.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim states cost reductions of '0.80%–2.02%' compared to accuracy-oriented methods. However, Table I reveals that this range is selectively reported. The full range of DF(Selector)–AO improvements across all eight forecasting settings and three selector variants is 0.11%–2.09%. Four of the eight settings (Parametric, Non-parametric, GAN-Separate, Diff-Separate) show improvements of only 0.11%–0.49%, well below the 0.80% lower bound claimed in the abstract. These are single-point estimates with no random-seed variation, no confidence intervals, and no significance testing reported for the cost comparisons. The paper mentions 10 runs only for timing benchmarks (§V.D, Fig. 12), not for the cost results in Table I. Without repeated trials and significance tests, it is unclear whether even the 0.80% improvements (VAE-Separate, GAN-Joint) are real or within run-to-run variance. Additionally, only a single 14-bus system with one dataset is tested, which is insufficient to support the generalizability claim across 'mainstream generative models.' The reader's KKT nonsingularity concern (Eq. 16) is mathematically valid but is a standard assumption in the OptNet literature [28] and is less directly load-bearing for the empirical claim than the absence of statistical validation. If the improvements for the settings reporting 0.11%–0.49% are within noise, the framework's benefit is limited to specific joint-forecasting configurations rather than being a general result.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"The paper proposes a unified decision-focused scenario generation and selection framework for distributionally robust optimization (DRO)-based power grid dispatch. The framework backpropagates operational cost gradients through the DRO optimization into generative models (VAE, GAN, diffusion), enabling scenario generation that is directly optimized for downstream dispatch cost rather than statistical accuracy. A differentiable scenario selector based on Gumbel-Softmax is introduced to reduce computational burden. Experiments on an IEEE 14-bus system with load data from southern China demonstrate cost reductions compared to accuracy-oriented methods.","tokens_in":20804,"tokens_out":1086,"duration_ms":220340,"significance":"The paper addresses a relevant problem at the intersection of decision-focused learning and power systems optimization. The model-agnostic design across VAE, GAN, and diffusion models is a strength, as is the differentiable scenario selector. The provision of a public code repository (https://github.com/hkuedl/Cost-oriented-Generative-Model) supports reproducibility. The framework is applied to a concrete two-stage Wasserstein-DRO dispatch formulation with tractable reformulation.","major_comments":[{"comment":"Table I and Abstract: The abstract claims cost reductions of '0.80%–2.02%' across different generative models. However, Table I shows that for several settings (Parametric, Non-parametric, GAN-Separate, Diff-Separate), the DF(Selector)-AO improvements range from 0.11% to 0.49%, well below the stated 0.80% lower bound. The abstract appears to selectively report only the better-performing settings. The authors should either revise the claimed range to reflect all results or explicitly state that the 0.80%–2.02% range applies specifically to the joint forecasting settings.","section":null},{"comment":"Table I: No statistical significance testing or random-seed variation is reported for the cost comparisons. The improvements in several settings are below 0.50% (e.g., 0.11% for Parametric with Selector-Inner), which may be within run-to-run variance. The paper mentions 10 runs only for timing benchmarks (§V.D, Fig. 12), not for the cost results. Without confidence intervals or repeated trials, it is unclear whether these small improvements are robust. The authors should provide multi-seed results with standard deviations or significance tests for the main cost comparisons.","section":null},{"comment":"§IV.A, Eq. (16): The KKT-based implicit differentiation assumes the Jacobian ∂K/∂x̃ is nonsingular at the optimum. The paper states this assumption but does not verify it for the specific DRO dispatch LP (12). LP degeneracy is common in power-system dispatch problems, and if the Jacobian is singular at degenerate vertices, the gradient computation is invalid. The authors should either provide conditions under which nonsingularity holds for the specific LP structure in (12) or discuss how degeneracy is handled in practice (e.g., via regularization or perturbation).","section":null}],"minor_comments":[{"comment":"§IV.B, Eq. (22): The smoothing coefficient ν is introduced but its value is not specified in the experimental setup. Please state the value used.","section":null},{"comment":"§IV.B, Eq. (23): The Gumbel-Softmax temperature τ_g is mentioned but its annealing schedule or final value is not provided. Please specify.","section":null},{"comment":"§IV.A, diffusion model backward propagation: The truncated backpropagation strategy uses n_tru steps, but the value of n_tru is not stated. Please include this in the experimental setup.","section":null},{"comment":"§V.A: The values of hyperparameters ε (ambiguity set radius), K (number of selected scenarios), and λ (regularization weight) are not clearly tabulated. A table summarizing all hyperparameter values would improve reproducibility.","section":null},{"comment":"Table I: The row label 'DF(Selector)-AO(Random)' is ambiguous. It is unclear whether this represents the best selector result minus AO, or a specific selector variant. Clarification is needed.","section":null},{"comment":"§V.C, Fig. 9: The x-axis values span 10^1 to 10^4, but the tick labels are not shown clearly. Please improve the readability of the axis labels.","section":null},{"comment":"§V.A: Only a single 14-bus system with one dataset is tested. While the authors mention the framework is applicable to other systems, testing on at least one additional system (e.g., IEEE 30-bus or 118-bus) would strengthen the generalizability claim.","section":null},{"comment":"The paper would benefit from a brief discussion of the computational overhead introduced by the OptNet backward pass relative to the forward solve, particularly for the full (non-simplified) system.","section":null}],"recommendation":"major_revision","confidential_remarks":"The core methodology is sound and the framework is well-motivated. The main concerns are empirical: the abstract overstates the results by selective reporting, and the absence of statistical validation for small improvements is a significant gap. The KKT nonsingularity concern is mathematically valid but is a standard assumption in the OptNet literature; a brief discussion or practical mitigation would suffice. If the authors address the statistical validation and correct the abstract's claims, the paper could be a solid contribution."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for the careful review and constructive feedback. The referee raises three major points: (1) the abstract's cost-reduction range appears to selectively report only better-performing settings; (2) no statistical significance testing or multi-seed variation is reported for the main cost comparisons; and (3) the KKT-based implicit differentiation assumes nonsingularity of the Jacobian, which may fail under LP degeneracy. We address each point below and commit to revisions where appropriate.","responses":[{"response":"The referee is correct that the abstract's stated range of 0.80%–2.02% does not encompass all settings reported in Table I. Examining the table, the improvements for the joint forecasting settings (VAE-Joint, GAN-Joint, Diff-Joint) range from 0.73% to 2.09%, while the separate forecasting and traditional methods show smaller improvements (0.11%–0.49%). The 0.80%–2.02% range was intended to characterize the joint forecasting results, where the decision-focused framework's benefits are most pronounced, but this scope was not made clear in the abstract. We will revise the abstract to explicitly state that the 0.80%–2.02% range applies to the joint forecasting settings, and we will also report the full range of improvements across all settings to avoid any impression of selective reporting.","revision_made":"yes","referee_comment":"Table I and Abstract: The abstract claims cost reductions of '0.80%–2.02%' across different generative models. However, Table I shows that for several settings (Parametric, Non-parametric, GAN-Separate, Diff-Separate), the DF(Selector)-AO improvements range from 0.11% to 0.49%, well below the stated 0.80% lower bound. The abstract appears to selectively report only the better-performing settings. The authors should either revise the claimed range to reflect all results or explicitly state that the 0.80%–2.02% range applies specifically to the joint forecasting settings."},{"response":"The referee raises a valid concern. We acknowledge that the cost results in Table I are reported as single-run values without confidence intervals or standard deviations, while the timing benchmarks in §V.D do use 10 runs. This is an inconsistency in our experimental reporting. We will conduct multi-seed experiments (at least 5–10 runs with different random seeds) for the main cost comparisons in Table I and report standard deviations. For settings where improvements are below 0.50%, we will explicitly discuss whether the differences are statistically significant. We note that for the joint forecasting settings where improvements exceed 0.73%, we expect these to be robust, but we will verify this empirically. We agree this is necessary for the reader to assess the reliability of the reported improvements.","revision_made":"yes","referee_comment":"Table I: No statistical significance testing or random-seed variation is reported for the cost comparisons. The improvements in several settings are below 0.50% (e.g., 0.11% for Parametric with Selector-Inner), which may be within run-to-run variance. The paper mentions 10 runs only for timing benchmarks (§V.D, Fig. 12), not for the cost results. Without confidence intervals or repeated trials, it is unclear whether these small improvements are robust. The authors should provide multi-seed results with standard deviations or significance tests for the main cost comparisons."},{"response":"The referee correctly identifies a technical gap. The nonsingularity of ∂K/∂x̃ at the KKT point is equivalent to the standard second-order sufficient condition (SOSC) for the LP, which requires linear independence of active constraints (LICQ) and strict complementarity. We acknowledge that LP degeneracy, where strict complementarity fails, is common in power-system dispatch problems and can arise in our formulation (12) when, for example, multiple generators hit their capacity limits simultaneously or when reserve constraints are binding at the same time step. We do not currently verify or handle this condition in practice. In the revised manuscript, we will add a discussion of this issue, including: (1) the specific conditions under which nonsingularity holds for the LP structure in (12), (2) a note that in our experiments we did not encounter numerical issues, which suggests degeneracy did not arise in the tested instances, and (3) a practical mitigation strategy such as adding a small quadratic regularization term to the LP objective or applying a perturbation to ensure strict complementarity. We note that such regularization is standard in the differentiable optimization literature (e.g., OptNet-based approaches). However, we cannot at this stage provide a formal proof that nonsingularity always holds for the specific DRO dispatch LP, so the discussion will be framed as conditions and practical mitigations rather than a guarantee.","revision_made":"partial","referee_comment":"§IV.A, Eq. (16): The KKT-based implicit differentiation assumes the Jacobian ∂K/∂x̃ is nonsingular at the optimum. The paper states this assumption but does not verify it for the specific DRO dispatch LP (12). LP degeneracy is common in power-system dispatch problems, and if the Jacobian is singular at degenerate vertices, the gradient computation is invalid. The authors should either provide conditions under which nonsingularity holds for the specific LP structure in (12) or discuss how degeneracy is handled in practice (e.g., via regularization or perturbation)."}],"tokens_in":20278,"tokens_out":1200,"duration_ms":166174,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"Bottom line: this paper does something genuinely useful — it wraps VAE, GAN, and diffusion models into a single decision-focused training pipeline for Wasserstein-DRO grid dispatch, with a differentiable Gumbel-Softmax scenario selector — but the empirical evidence is thinner than the abstract suggests and needs hardening before publication. The code is public, which matters. The framework design is clean: the model-agnostic abstraction over three generative paradigms is legitimate, the OptNet backward pass through the DRO LP is correctly derived, and the correlation analysis (Figs. 7-8) convincingly shows that joint generation captures inter-bus dependence that separate forecasting misses. The scenario selector comparison against K-means, K-medoids, and hierarchical clustering (Table II) is a fair benchmark and the proposed selector wins consistently. That said, two problems stand out. First, the abstract claims 0.80–2.02% cost reduction, but Table I shows four of eight settings at 0.11–0.49%. The abstract is cherry-picking the better results. Second, there are no confidence intervals, no random-seed variation, and no significance tests on the cost comparisons. The 10-run protocol is used only for timing benchmarks, not for the cost numbers that actually matter. For improvements in the 0.1–0.5% range, single-point estimates are not enough. On the KKT nonsingularity concern (Eq. 16): this is a standard assumption in the OptNet literature and the paper cites Amos & Kolter correctly. It would be better to verify it for this specific LP structure, but I would not call it a load-bearing flaw — the OptNet machinery is well-established and the empirical results are consistent with the gradients being computed correctly. The bigger issue is scale: one 14-bus system, one dataset. The generalizability claim across 'mainstream generative models' needs at least one larger test system. The truncated backpropagation for diffusion (n_tru steps) is a reasonable engineering choice but the hyperparameter is not reported. This paper is for researchers working at the intersection of decision-focused learning and power system optimization. It deserves a serious referee who can ask for: (1) honest reporting of the full improvement range, (2) repeated trials with significance tests on Table I, (3) at least one larger system, and (4) reported hyperparameters. The framework is sound enough to warrant the revision cycle.","headline":"Unified decision-focused scenario generation for DRO dispatch: real framework, selectively reported results","tokens_in":21583,"tokens_out":565,"would_cite":true,"duration_ms":177442,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Train scenarios on grid cost, not forecast error: 0.8–2% savings","keywords":["decision-focused learning","scenario generation","distributionally robust optimization","power system dispatch","generative models","scenario selection","Wasserstein ambiguity set","KKT differentiation"],"falsifier":"If the DRO dispatch LP frequently encounters degenerate vertices (where the KKT Jacobian is singular), the gradient computation in Equation 16 becomes invalid, and the decision-focused training signal is corrupted. One could test this by tracking the condition number of the KKT Jacobian across training iterations on realistic grid data and checking whether degenerate points arise often enough to destabilize training.","tokens_in":20762,"feed_emoji":"⚡","tokens_out":1208,"duration_ms":228929,"temperature":0.7,"pith_summary":"This paper argues that the scenarios fed into distributionally robust optimization (DRO) for power grid dispatch should be trained to minimize downstream operational cost, not statistical forecasting error. The authors build a unified framework that wraps three major generative model families—variational autoencoders (VAEs), generative adversarial networks (GANs), and diffusion models—into a single decision-focused training loop. The key mechanism is a differentiable pipeline: scenarios are generated from a learned joint distribution that captures inter-bus correlations, passed through a differentiable scenario selector that reduces the scenario count for tractability, and then fed into a DRO dispatch formulation whose optimal solution is differentiated via its KKT conditions using the OptNet approach. The gradient of the resulting operational cost flows back through the selector and into the generative model, updating its parameters so that future scenarios are more decision-relevant. The paper demonstrates that this approach reduces operational cost by 0.80%–2.02% across all three generative model types compared to accuracy-oriented training, and that the learned scenario selector consistently outperforms statistical clustering methods like K-means and hierarchical clustering.","feed_headline":"Train scenarios on grid cost, not forecast error: 0.8–2% savings","feed_subtitle":"A unified framework turns VAEs, GANs, and diffusion models into decision-focused scenario generators for robust grid dispatch, cutting cost.","key_machinery":"The framework has three mechanical components. First, a generative model (VAE decoder, GAN generator, or diffusion denoising network) produces a pool of jointly correlated scenarios from a shared latent space. Second, a differentiable scenario selector parameterized by a neural network outputs soft selection weights via Gumbel-Softmax reparameterization, constructing a reduced scenario set as convex combinations of pool scenarios during training and performing hard discrete selection at evaluation. Third, the DRO dispatch problem is reformulated as a linear program whose KKT conditions are differentiated using OptNet, yielding the sensitivity of the optimal dispatch cost to the inputscenario","core_discovery":"The central discovery is that a single decision-focused training protocol can be applied across architecturally distinct generative models (VAE, GAN, diffusion) for scenario generation in DRO-based grid dispatch, by abstracting each model into a common forward interface and routing the operational-cost gradient back through model-specific backward paths. The protocol works because the DRO dispatch problem, when reformulated as a linear program, admits differentiation through its KKT conditions, creating a valid gradient path from dispatch cost to generative model parameters. A secondary discovery is that a learned, differentiable scenario selector trained under the same cost objective can be","pith_inferences":["The 0.80%–2.02% cost reduction is measured on a single 14-bus system with loads from one region of China. Whether the relative benefit scales to larger transmission networks with hundreds of buses and more complex topology remains an open question—the computational cost of the KKT-based backward pass grows with problem size, and the benefit may be diluted or amplified in larger systems.","The framework's success hinges on the DRO dispatch LP having a locally unique KKT point with a nonsingular Jacobian. Power system dispatch problems frequently encounter degenerate LP vertices where multiple constraints bind simultaneously, which would corrupt the gradient signal. The paper does not report how often this occurs in practice or how the training behaves when it does.","The truncated backpropagation strategy for diffusion models (keeping only the last n_tru denoising steps) introduces an approximation whose quality likely depends on the variance schedule and the value of n_tru. The trade-off between gradient fidelity and memory cost is not fully characterized, and the optimal truncation depth may vary with the complexity of the uncertainty distribution."],"forward_implications":["Power system operators could adopt decision-focused scenario generation as a drop-in replacement for accuracy-oriented forecasting pipelines, obtaining dispatch decisions that directly account for the asymmetric cost structure of over- vs. under-forecasting.","The model-agnostic design means that as new generative architectures emerge (e.g., flow matching, transformer-based generators), they can be integrated into the same decision-focused training loop without redesigning the differentiation pipeline.","The differentiable scenario selector could be applied to other computationally expensive optimization problems beyond grid dispatch, such as supply chain or portfolio optimization, where scenario reduction is needed but statistical similarity does not capture decision relevance.","Joint distribution modeling across buses eliminates physically implausible scenario combinations, potentially improving grid reliability metrics that are not captured by cost alone."],"fun_headline_variants":["Cut grid dispatch cost by up to 2% with decision-focused scenarios","Train generative models on dispatch cost, not forecast accuracy","Cost-driven scenario generation lowers robust dispatch cost","Differentiable scenario selection trims grid dispatch cost","Optimize grid scenarios for dispatch cost to save up to 2%"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The gradient computation that enables the entire decision-focused training loop assumes that the DRO dispatch linear program has a locally unique KKT point and that the Jacobian of the KKT system is nonsingular at the optimum. The paper states this assumption but does not verify it for the specific dispatch formulation used, and degenerate LP vertices—where multiple constraints bind simultaneously—are common in power system dispatch problems and would make the Jacobian sing","fun_headline_variants_meta":{"raw":{"variants":["Cut grid dispatch cost by up to 2% with decision-focused scenarios","Train generative models on dispatch cost, not forecast accuracy","Cost-driven scenario generation lowers robust dispatch cost","Differentiable scenario selection trims grid dispatch cost","Optimize grid scenarios for dispatch cost to save up to 2%","Unifying VAEs, GANs, and diffusion for cost-aware dispatch","Scenario generation optimized for grid dispatch cost"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":1247,"prompt_tokens":522,"completion_tokens":725,"prompt_tokens_details":null},"tokens_in":522,"tokens_out":725,"duration_ms":37658,"temperature":1.0,"reasoning_tokens":673,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T22:51:50.617358+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If the DRO dispatch LP frequently encounters degenerate vertices (where the KKT Jacobian is singular), the gradient computation in Equation 16 becomes invalid, and the decision-focused training signal is corrupted. One could test this by tracking the condition number of the KKT Jacobian across training iterations on realistic grid data and checking whether degenerate points arise often enough to destabilize training.","supporting_citations":[],"review_version":1}