{"id":"1c7e21a6-6038-4c99-8afa-90914a4e43d2","arxiv_id":"2607.17514","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Jointly training workload prediction with downstream grid scheduling as a differentiable pipeline reduces simulated cost and capacity violations versus a two-stage baseline, at the price of much larger forecast error.","lead":"An AI data center's workload predictor and its power-grid scheduler are trained together in one loop, so forecasts are tuned for cheap, safe scheduling instead of raw accuracy. On a simulated 33-bus grid, this cuts reported operating cost and capacity violations versus a two-stage train-then-schedule baseline, while deliberately accepting higher forecast error.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing workload-to-capacity scaling in Eq. (9) makes the optimization infeasible as written and undermines the reported cost/violation comparison.","rationale":"The paper proposes a plausible end-to-end decision-focused learning framework, and the mathematical machinery (implicit differentiation of convex programs) is standard and coherent. However, the central empirical claim—that PTS significantly reduces operational cost and improves constraint satisfaction—rests entirely on the numerical experiments in Section IV. The most load-bearing flaw is the internal inconsistency between the workload normalization to [0,1], the capacity P_full=0.15, and Eq. (9), which uses the normalized workload directly without any scaling factor. This makes the decision-layer optimization (10) infeasible for typical inputs, meaning the experiments as described cannot have run without an unstated rescaling or softening of constraints. That rescaling would directly determine the relative magnitudes of generation cost and shedding penalty, thereby controlling the entire cost/violation comparison in Table I. The reader identified exactly this as the weakest assumption; I agree. Other concerns (shedding penalty conflated with cost, lack of seeds/error bars, no code) are secondary and addressable, but the scaling issue is the one that must be settled before any of the reported numbers can be trusted. A concrete check—recomputing feasibility and re-running with an explicit scaling factor—would decisively resolve whether the qualitative conclusion survives. Since the reader already set the verdict to CONDITIONAL and my concern aligns with that, the verdict should remain unchanged: the paper is conditionally accepted pending clarification and re-validation of the experimental setup.","tokens_in":11864,"tokens_out":7073,"duration_ms":61935,"concrete_test":"Compute the maximum feasible total workload for Problem (10) under the stated equations: with T=24, one datacenter, P_idle=0, P_full=0.15, the total scheduled energy cannot exceed 24×0.15=3.6. Sum the test-set workloads used in Table I; if the sum exceeds 3.6, (10) is infeasible. Then ask the authors to report the scaling factor applied to w in Eq. (9), or re-run the experiments with w scaled by α=P_full/max(w) so that the decision problem is feasible, and compare the resulting Table I cost and violation metrics. If the qualitative advantage of PTS disappears, the original claim was an artifact of the missing scaling.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV states the workload is normalized to [0,1] (IV.A.2) and the datacenter capacity P_full=0.15 units (IV.A.1). Equation (9) sets P_IT_i,t = Σ_j Σ_s ŵ_j,s M_j,i,s,t + P_idle, with M constrained by (8b) so that all arriving workload is scheduled within the horizon. With one datacenter and P_idle=0, if ŵ_j,s can be as high as 1, the total IT energy over T=24 equals Σ_s ŵ_s, so the average P_IT equals the average ŵ; a value of 1 at any hour forces P_IT=1 > 0.15. The hard capacity constraint (6) in the decision layer (10) therefore makes the optimization infeasible for typical normalized workloads. The only way the reported experiments could run is an unstated rescaling of w (e.g., multiplying by a factor ≤0.15) or softening the capacity constraint in the decision layer; neither appears in the paper. This rescaling changes the relative weight of generation cost versus shedding penalty in the over-shifted loss (13), so the numerical comparison in Table I (cost loss 10.72 vs 31.84, violation 0.0007 vs 0.7518) is not reproducible or interpretable. The claimed advantage of PTS may be an artifact of this unit mapping, and the learned over/under-forecasting described in §IV.B.2 is a direct consequence of it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an end-to-end Predict-then-Schedule (PTS) framework for coordinating delay-tolerant AI data center workload with distribution network operation. An LSTM workload predictor is trained jointly with a differentiable convex optimization layer that solves a DistFlow-based grid scheduling problem. A decision-focused loss, called the over-shifted loss, combines electricity generation cost with a quadratic penalty on load shedding. Simulations on a modified IEEE 33-bus system with a single data center compare PTS against a conventional two-stage predictor and a reserve-based baseline. The reported results show that PTS achieves much lower cost loss and constraint violations at the expense of higher prediction MSE. The central claim is that end-to-end decision-focused training significantly reduces operational cost and enhances security relative to the conventional two-stage baseline.","tokens_in":12193,"tokens_out":5829,"duration_ms":51931,"significance":"The paper addresses a timely and important problem: coordinating AI data center workload with distribution grid constraints through decision-focused learning. The mathematical framework is standard, and the use of implicit differentiation through a convex optimization layer is appropriate. If the empirical claims were properly supported, the work would offer a useful demonstration that end-to-end training can outperform a two-stage pipeline for this application. The formulation of the decision layer and the evaluation layer is clear, and the idea of an over-shifted loss to penalize load shedding is reasonable. However, the current experimental evidence has major gaps: the workload-capacity scaling is not specified, the reported cost metric mixes a tunable penalty with actual electricity cost, the violation metric is undefined, and no code, data, or error bars are provided. These issues prevent the central claim from being verified as presented.","major_comments":[{"comment":"The simulation setup is internally inconsistent. In IV.A.1 the data center full power capacity is P_full=0.15 units, while IV.A.2 states the workload is normalized to [0,1]. Equation (9) sets P_IT_i,t = sum_j sum_s \\hat w_{j,s} M_{j,i,s,t} + P_idle_i. With one data center and P_idle=0, any hour with \\hat w > 0.15 violates the hard capacity constraint (6). The decision layer (10) enforces (6), so the optimization is infeasible for typical normalized workloads unless an unstated rescaling is applied. The paper does not specify any scaling factor between the workload data and the power capacity, nor any relaxation of the decision-layer capacity constraint. Since the relative magnitude of generation cost versus shedding penalty in the over-shifted loss (13) depends directly on this mapping, the numerical results in Table I are not reproducible or interpretable. Please clarify the unit mappin","section":"Section IV.A, Eqs. (6)-(9)"},{"comment":"The over-shifted loss L_cost in Eq. (13a) is the sum of electricity generation cost and a quadratic penalty gamma (P_shed)^2. This penalty has units that do not match the generation cost, and gamma is a user-chosen weight. Table I reports 'Cost Loss' using this composite quantity, so the comparison between PTS and Two-Stage conflates actual operational cost with an arbitrary penalty term. For the Two-Stage baseline, the high cost loss is largely penalty from violations (violation 0.7518), not generation cost. The abstract's claim of 'significantly reduces operational cost' is not substantiated unless the electricity cost and the shedding penalty are reported separately, or the penalty is given a physically motivated monetary interpretation. Please report the pure generation cost component and justify the choice of gamma.","section":"Section III-B, Eq. (13), Table I"},{"comment":"The metric 'Violation' appearing in Table I and Fig. 3b is never defined. The text refers to 'magnitude of constraint violations' but gives no formula, units, or aggregation method. Since the security claim rests on the comparison 0.0007 vs 0.7518, the definition is essential. Please specify whether this is total load shedding, the maximum per-bus voltage or capacity violation, an average over the horizon, or another quantity, and provide its units.","section":"Section IV.B, Table I, Fig. 3b"},{"comment":"The experimental evaluation reports a single run with no error bars, no multiple seeds, and no code or data release. The workload dataset is described as 'synthesized from real-world power demand forecasting data', but the source and preprocessing steps are not given. Given that the central claim is an empirical comparison, the absence of variance information and reproducibility artifacts makes it impossible to assess the statistical significance of the reported differences. Please provide code/data (or a detailed description), report results over multiple seeds, and include confidence intervals or standard deviations.","section":"Section IV.A, Table I"}],"minor_comments":[{"comment":"The voltage constraint in Eq. (4) appears to have missing superscript bars: it is written as 'v 2 i ≤V i,t ≤ v2 i ', presumably meaning \\underline{v}_i^2 and \\overline{v}_i^2. Please correct the typesetting.","section":"Eq. (4)"},{"comment":"The captions and axis labels of Figures 4 and 5 contain garbled text due to font encoding issues. They should be regenerated so the figure content is legible.","section":"Fig. 4 and Fig. 5"},{"comment":"The statement 'A sufficiently large \\gamma prioritizes the elimination of load shedding' is correct but the sensitivity of the results to \\gamma is only shown for two values (10 and 100). A brief discussion of how \\gamma is chosen in practice would help.","section":"Section III-B"}],"recommendation":"major_revision","confidential_remarks":"The workload-capacity scaling issue is the most serious concern. If the authors cannot provide the missing mapping between the normalized workload data and the data center capacity, the central comparison may be invalid. I would encourage the editor to request the code and data as part of the revision to verify the reported numbers. The framework itself is sound and potentially publishable after the empirical validation is clarified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read of arXiv:2607.17514. The paper takes the standard predict-then-optimize machinery (Donti, Elmachtoub, Amos, Agrawal) and applies it to coordinating AI data center workload scheduling with a distribution grid. That's a timely and sensible thing to do, and the authors get the core mechanism right: train the forecaster against the downstream scheduling loss, use implicit differentiation through a convex optimization layer, and let the predictor deliberately distort forecasts to avoid constraint violations. The over-shifted loss with a load-shedding penalty is a reasonable way to keep the evaluation problem feasible and differentiable. If I had a student working on data center demand response, I'd point them here.\n\nThe math is standard but coherent. The empirical story is the expected one: PTS trades MSE for much lower constraint violations and a lower combined cost-vs-penalty loss than two-stage. The learned over-forecasting during cheap hours and under-forecasting during peak hours is a nice illustration of decision-focused behavior.\n\nThe soft spots are real and mostly in the experiments. The biggest one: Section IV says the workload is normalized to [0,1] and the data center capacity P_full is 0.15 units. With Eq. (9) and the conservation constraint (8b), a workload series with a mean above 0.15 will force P_IT above P_full at some hour regardless of scheduling. Since typical normalized load data has a mean well above 0.15, the optimization would be infeasible as written. The authors must have rescaled the workload or softened the constraint; they don't say which, and that rescaling changes the relative weight of generation cost vs shedding penalty in the reported loss. So the Table I numbers are not reproducible or interpretable. Also, 'Cost Loss' includes the load-shedding penalty, not a pure electricity cost, so 'reduces operational cost' is overstated if that's all they report. There are no error bars, no multiple seeds, and no code/data. These are all fixable, but they need to be fixed before the numbers can be taken at face value.\n\nWho is this for? Researchers working on decision-focused learning for power systems or data center-grid coordination. It's a useful case study, not a breakthrough. I'd send it to peer review, but with a clear request for the scaling details, a separate electricity-cost metric, and code/data. The qualitative conclusion is plausible and consistent with prior PTO results, so I wouldn't reject it.","headline":"Useful decision-focused learning application with a genuine reproducibility gap in the workload-capacity scaling.","tokens_in":12705,"tokens_out":3379,"would_cite":false,"duration_ms":30518,"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 proposes an end-to-end Predict-then-Schedule framework that trains the workload predictor through the downstream scheduling optimization, so the model learns forecasts that reduce operational cost and grid violations rather than","keywords":["predict-then-optimize","decision-focused learning","AI data center","workload scheduling","power distribution network","differentiable convex optimization","load shedding","implicit differentiation"],"falsifier":"Rerun the experiments with explicit physical units and a stated peak-workload-to-capacity ratio, and check whether PTS still beats the two-stage baseline and still learns to over- and under-forecast; if the advantage vanishes or the learned distortion reverses, the central claim depends on the unstated rescaling.","tokens_in":11732,"feed_emoji":"⚡","tokens_out":4302,"duration_ms":37858,"temperature":0.7,"pith_summary":"The paper argues that for AI data center workloads that can be shifted in time and space, training the upstream predictor to minimize the downstream scheduling cost—rather than forecast error—yields schedules that are cheaper and safer. It builds a three-layer pipeline: an LSTM predictor, a differentiable convex optimization layer that solves the grid-constrained scheduling problem, and an evaluation layer that penalizes load shedding via an over-shifted loss. On a modified IEEE 33-bus system, the approach reports a substantial reduction in operational cost and near-zero constraint violations, while deliberately accepting higher mean squared error. The point is that statistical accuracy is not the right objective when predictions feed an optimizer; decision quality is what matters. If true, this supports decision-focused learning for grid-coordinated data center operations.","feed_headline":"End-to-end training cuts AI datacenter grid costs","feed_subtitle":"Coupling workload prediction with grid-aware scheduling cuts cost and violations, at the price of forecast accuracy.","key_machinery":"The key mechanism is the differentiable convex optimization layer: the downstream scheduling problem is written in compact form with objective J(y,ϕ), linear equalities and inequalities, and its KKT conditions are differentiated implicitly to obtain dy*/dϕ. This lets the gradient of the over-shifted loss L_cost flow back through the optimizer into the prediction network parameters θ. The over-shifted loss itself is the second load-bearing piece: it relaxes the hard capacity constraint with a load-shedding variable P_shed and adds a quadratic penalty γ(P_shed)^2, so the evaluation problem remains feasible and gradients stay continuous while the model is incentivized to hedge against violation","core_discovery":"The central claim is that end-to-end decision-focused training is superior to the conventional two-stage pipeline for coordinating delay-tolerant AI workloads with distribution network constraints. The framework treats the convex scheduling problem as a differentiable layer using implicit differentiation through the KKT system, so the gradient of an over-shifted loss—which combines generation cost and a quadratic penalty for load shedding—can be backpropagated into the predictor. The learned forecasts intentionally deviate from ground truth, over-forecasting at some hours and under-forecasting at others, to steer the optimizer away from infeasible, high-cost decisions. With penalty coefficie","pith_inferences":["The reported unit mismatch—workload normalized to [0,1] versus a datacenter capacity of 0.15 units—suggests the comparison depends on an unstated rescaling of the workload data; if the data were not rescaled, the scheduling problem would be infeasible without massive shedding, so the learned hedging behavior may be partly an artifact of that scaling choice.","A natural testable extension is to rerun the comparison with a two-stage baseline that uses an optimized safety margin rather than a fixed 80% reserve; the advantage of PTS might shrink or disappear under that stronger baseline.","The implicit differentiation assumes LICQ and strict complementarity; in degenerate cases such as load shedding at zero or binding bounds, the gradient estimates may be biased, so robustness to degeneracy is worth probing."],"forward_implications":["Because the predictor is trained through the optimizer, the model learns to hedge: it deliberately distorts forecasts (over-forecasting at low-price hours, under-forecasting at peak) to avoid infeasible schedules.","The over-shifted loss with a sufficiently large γ drives violations to near zero; both γ=10 and γ=100 converge to similar cost levels, so the framework is robust to the penalty magnitude.","A fixed 80% capacity reserve reduces violations but is suboptimal in cost compared to learned hedging, implying that decision-aware training can replace static safety margins.","Because the downstream problem remains an exact convex program, the scheduling decisions stay interpretable while training end-to-end."],"fun_headline_variants":["End-to-end AI grid scheduling cuts costs","Predict-then-schedule beats two-stage for AI loads","Decision-focused training trims datacenter power costs","AI datacenters: direct scheduling optimization slashes costs","End-to-end PTS framework reduces grid strain and spend"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the workload data, normalized to [0,1], can be fed into a scheduling model whose full capacity is 0.15 units; without an unstated rescaling, the optimization would be infeasible, so the reported cost and violation comparison rests on that hidden unit mapping.","fun_headline_variants_meta":{"raw":{"variants":["End-to-end AI grid scheduling cuts costs","Predict-then-schedule beats two-stage for AI loads","Decision-focused training trims datacenter power costs","AI datacenters: direct scheduling optimization slashes costs","End-to-end PTS framework reduces grid strain and spend"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000141,"raw_usage":{"total_tokens":961,"prompt_tokens":667,"completion_tokens":294,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":411,"completion_tokens_details":{"reasoning_tokens":217}},"tokens_in":411,"tokens_out":294,"duration_ms":3066,"temperature":1.0,"reasoning_tokens":217,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T17:44:27.710301+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the experiments with explicit physical units and a stated peak-workload-to-capacity ratio, and check whether PTS still beats the two-stage baseline and still learns to over- and under-forecast; if the advantage vanishes or the learned distortion reverses, the central claim depends on the unstated rescaling.","supporting_citations":[],"review_version":1}