{"id":"a82d37d0-27eb-472f-b952-3d9207467cd9","arxiv_id":"2501.03443","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Neural proxies with repair and completion layers can learn to solve parametric optimization problems in milliseconds, returning feasible solutions with empirically tight dual bounds on large power-grid instances.","lead":"This paper presents a unifying framework, \"optimization learning,\" for training neural-network proxies that map optimization problem inputs to feasible, near-optimal solutions, combining deep learning with repair or completion layers. It reviews three proxy designs (primal, dual, primal-dual) and reports order-of-magnitude speedups on power-system benchmarks such as economic dispatch and security-constrained optimal power flow.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'trustworthy by design' claim fails for the flagship PDL-SCOPF application: Section 7.3 admits the binary-search layer can violate contingency power balance, making feasibility a training outcome, not an architectural guarantee.","rationale":"The reader identified the repair-layer requirement and explicitly cited Section 7.3; I agree with that direction but would place more weight on the PDL-SCOPF contingency-balance admission because it directly contradicts the abstract's unconditional feasibility claim for the flagship application. This is an internal inconsistency, not a disagreement with external consensus: Section 5.2's E2ELR repair layers for economic dispatch are constructed to be feasible when a feasible solution exists, and the empirical results are strong. But the paper's own open problem in Section 8 admits that general repair-layer design is unsolved, and Section 7.3 admits that PDL-SCOPF's contingency balance is only satisfied when training completes in experiments. The proposed audit would show whether a fully trained model is empirically robust, but no audit can convert an empirical property into a design guarantee. The right outcome is therefore the same conditional verdict: the paper should qualify the by-design language and add the missing worst-case and distribution-shift analysis before it can support the headline claim.","tokens_in":20388,"tokens_out":6455,"duration_ms":65312,"concrete_test":"Audit PDL-SCOPF on the 6515 rte case by checkpointing the trained primal network at outer iterations T=5, 10, 15, 20 of Algorithm 1 and evaluating each checkpoint on 10,000 held-out instances from the Section 7.4 distribution plus a stress set with the load, cost, and upper-bound perturbation factors multiplied by 1.5. For every generator contingency, compute e_k = 1^T p_k - 1^T d and report the maximum and 99.9th percentile of |e_k| over all contingencies and instances, with the 1e-4 p.u. tolerance used in Tables 11-12. If any stress-set or intermediate-checkpoint instance exceeds that tolerance, feasibility is a training artifact, not a by-design guarantee. Also report variance and worst-case optimality gaps, since Tables 11-13 report only maxima or means over 1,000 instances.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim requires feasibility to be a by-design property of the optimization proxy. The paper's own Section 7.3 states: 'Contrary to its CCGA counterpart, Algorithm 2 may not always satisfy the power balance constraints in the contingencies for reasons described in [37]; They are however satisfied when training completes in the experiments.' PDL-SCOPF is one of the two headline applications and the basis of the 'four orders of magnitude faster' conclusion. Its contingency power-balance constraints are therefore not enforced by a repair layer; they are an empirical property of a converged training run. If training is stopped early, uses a different seed or architecture, or encounters instances outside the training distribution, those constraints can be violated. The abstract offers no such caveat. The same limitation extends to the general methodology: Section 8 lists 'understanding how to derive effective repair layers for a wide range of applications' as an open problem. In addition, no certified quality guarantee is provided for PDL-SCOPF; the reported optimality gaps are empirical. Thus the strongest form of the claim, feasible and quality-guaranteed by design, is not supported by the demonstration.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces 'optimization learning' as a framework for learning parametric optimization mappings with differentiable programs that combine neural networks with repair or completion layers. Three proxy architectures are presented: primal optimization proxies, dual optimization proxies, and primal-dual learning. The framework is illustrated on economic dispatch with reserve requirements, DC optimal power flow, and security-constrained optimal power flow, using PGLib benchmarks with up to 30,000 buses and reporting millisecond inference times and speedups of several orders of magnitude over commercial solvers. The central claim is that these proxies are 'trustworthy by design': they compute feasible solutions, provide quality guarantees, and scale to large instances.","tokens_in":20596,"tokens_out":5829,"duration_ms":55853,"significance":"If the claims were fully supported, the framework would be practically significant for real-time power-system operations: self-supervised training avoids expensive optimal labels, inference is orders of magnitude faster than state-of-the-art solvers, and the repair/completion layers are intended to make feasibility an architectural property. The derivations shown are largely sound: the balance repair layer in Eq. (5) is a differentiable feasible-point construction, the dual completion layer Clp produces dual-feasible points by construction, and the ALM-style primal-dual updates in Eqs. (12)-(13) are coherent. The experimental comparisons against Gurobi, Mosek, and CCGA on standard benchmarks are concrete and reproducible in spirit. However, the 'trustworthy by design' claim is not fully supported: the flagship PDL-SCOPF application does not have a by-construction guarantee for the contingency power-balance constraints, and the reported optimality gaps are empirical rather than certified. The contribution is therefore valuable as a synthesis and demonstration, but its stated guarantees need to be qualified.","major_comments":[{"comment":"The abstract and Section 8 state that optimization proxies are 'trustworthy by design' because they compute feasible solutions. For PDL-SCOPF, the text admits that Algorithm 2 'may not always satisfy the power balance constraints in the contingencies' and that these constraints are satisfied only 'when training completes in the experiments.' This makes feasibility an empirical property of a converged training run, not an architectural guarantee. Since PDL-SCOPF is one of the two headline applications and underlies the four-orders-of-magnitude speedup claim in Section 7.4, the central claim is overstated. The paper should either soften the 'by design' language to 'empirically feasible on the tested instances' or add a certified repair step for the contingency power-balance constraints.","section":"§7.3, Algorithm 2"},{"comment":"The paper claims that optimization proxies 'provide quality guarantees on returned solutions.' The dual completion layer in Section 6 does construct dual-feasible points, hence valid lower bounds, by construction. However, the optimality gaps reported for the primal proxies in Tables 2 and 13 are empirical means over test sets, and Table 7 reports empirical dual-gap ratios; no certified primal-dual gap is produced for the applications. The phrase 'quality guarantees' should therefore be replaced by 'empirical quality' for primal solutions, or the paper should demonstrate a principled primal-dual gap procedure that ties the dual bound to the returned primal solution.","section":"§5.4.3, §6, §7.4, Table 13"},{"comment":"Section 8 lists 'understanding how to derive effective repair layers for a wide range of applications' as an open problem. The general methodology therefore depends on an application-specific design step whose tractability is not established. This limitation is acknowledged in the conclusion, but it should also be reflected in the abstract: the 'by design' guarantee is conditional on the existence of an efficient repair or completion layer for the problem class. Stating this as a formal assumption at the outset, rather than as an open problem at the end, would make the scope of the claim precise.","section":"§8"}],"minor_comments":[{"comment":"The subscript in the multiplier update contains a spurious '<' character: 'Mθt<+1(xi)' should be 'Mθ_{t+1}(xi)'.","section":"§4.2"},{"comment":"There are several typographical errors, including 'witout', 'imlpicit', 'opimization', 'PDF-SCOPF' for PDL-SCOPF, and 'strenghts'; a careful proofread is needed.","section":"§5.2, §7.4, §8"},{"comment":"The sentence 'As shown in Figure 2' in the description of the PDL-SCOPF primal network is misleading; the intended reference is likely Figure 10.","section":"§7.3"},{"comment":"The statement that Algorithm 2 may fail 'for reasons described in [37]' leaves a load-bearing explanation in a cited paper; a one-sentence summary of those reasons would make the limitation self-contained.","section":"§7.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is essentially a condensed synthesis of the author's own prior papers [9, 23, 36, 37], with no new experiments or derivations beyond the unification. The high self-citation density is understandable for a research-program survey and should not by itself count against the paper. The main issue is that the abstract and conclusion assert guarantees that the flagship application does not deliver. If the editor values the synthesis, the paper can become acceptable once the overclaims are corrected and the limitations stated in Section 8 are incorporated into the framing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a read if you want a compact map of the optimization-learning area, but don't take the abstract at face value. The paper is essentially a keynote-style review that unifies a decade of Van Hentenryck's group work on primal, dual, and primal-dual proxies. That framing—feasible primal proxies via repair layers, dual proxies via completion layers, and ALM-style primal-dual learning—is genuinely useful and is the real contribution here. The derivations are correct: the balance repair layer exactly restores equality, the dual completion Clp(qz) = (qz, |cx - qz Ax|+, |cx - qz Ax|-) produces dual-feasible points by construction, and the PDL updates follow standard augmented Lagrangian logic. The experimental chapters are honest summaries of work validated against Gurobi, Mosek, and CCGA, and the speed numbers are striking.\n\nThe soft spot is sizing. The abstract says the proxies are 'trustworthy by design: they compute feasible solutions... provide quality guarantees.' That is true for some parts and not for others. For the ED primal proxy, feasibility is architectural. For the dual LP proxy, dual feasibility is architectural. But optimality/quality is empirical in every case—there is no certificate, just tight observed gaps. And in the flagship PDL-SCOPF application, Section 7.3 explicitly says the binary-search layer 'may not always satisfy the power balance constraints in the contingencies' and that they are satisfied only when training completes. The stress-test note has this right. Feasibility there is a training outcome, not a by-design guarantee. The paper also lists 'how to derive effective repair layers for a wide range of applications' as an open problem in Section 8, which undercuts the general claim further. These are framing problems, not fatal ones: the body is candid about the limitations, and the methods themselves look solid.\n\nOne more thing: reproducibility. The experimental tables are mostly pulled from earlier papers, and variance is not always reported. That is acceptable for a review, but it makes the paper a pointer to the literature rather than a self-contained record. The heavy self-citation is not a flaw here—it is an honest description of the paper's role.\n\nBottom line: this is a good synthesis that deserves serious refereeing, but the abstract and conclusion need to be pulled back to match what is actually shown. The audience is people who want an entry point into constrained optimization learning, especially in power systems. Send it to review, but insist the 'trustworthy by design' language be scoped.","headline":"A clear, well-written synthesis of the author's own optimization-learning line, but the 'trustworthy by design' claim overreaches: the flagship SCOPF application gets feasibility from training, not architecture.","tokens_in":21188,"tokens_out":3987,"would_cite":true,"duration_ms":31160,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","90C31"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that optimization proxies—differentiable programs that combine a neural network with repair or completion layers—can learn the input-output map of a parametric optimization problem and return feasible, near-optimal…","keywords":["optimization learning","optimization proxies","parametric optimization","self-supervised learning","primal-dual learning","economic dispatch","security-constrained optimal power flow","differentiable programming"],"falsifier":"On a random batch of held-out instances from the same load distribution, evaluate the trained primal and dual proxies on the same grid; the central claim fails if any primal output violates a hard constraint, or if the primal objective falls below the dual objective, since that would mean the dual proxy is not a valid lower bound.","tokens_in":20133,"feed_emoji":"⚡","tokens_out":12615,"duration_ms":111309,"temperature":0.7,"pith_summary":"The paper is trying to establish that the repeated solving of a parametric optimization problem can be replaced by a trained differentiable program, called an optimization proxy, that outputs feasible points, while a companion dual proxy outputs a valid lower bound. This matters because many engineering optimization problems are solved over and over on nearly identical inputs, and exact solvers are too slow to run in real time or inside an interactive decision loop. If the proxy behaves as claimed, the same problem class can be answered in milliseconds with a certified bound, and end-to-end self-supervised training removes the need for precomputed optimal labels. The paper demonstrates the concept on power-system applications: economic dispatch for grids up to 30,000 buses, DC optimal power flow, and security-constrained optimal power flow under N-1 contingencies.","feed_headline":"Power-grid optimization solved 10,000x faster with learned proxies","feed_subtitle":"Neural nets plus repair layers produce feasible, near-optimal solutions in milliseconds, backed by a dual lower bound.","key_machinery":"The machinery is the repair/completion layer: a hard-coded, differentiable map that turns whatever the neural network predicts into a point that satisfies the problem's constraints. The generic template is the closest-point projection, $\\min_y \\|y - p_y\\|$ subject to $h_x(y)=0$ and $g_x(y)\\ge 0$, but the paper's practical examples are hand-derived closed forms: proportional rescaling for power balance and reserves in economic dispatch, positive/negative-part completion for linear-programming duals, and a binary-search layer for contingency dispatches in security-constrained OPF. Because each layer has subgradients almost everywhere, the original objective can be backpropagated through the whole model, making the composed system a differentiable program that is also fast at inference.","core_discovery":"The central claim is that parametric optimization can be outsourced to a proxy that is trustworthy by design: a trained composition of a neural network and a repair or completion layer is constructed to land in the feasible set, and a dual proxy supplies a certified lower bound. Unlike a regression model, the proxy is not merely close to a solution; its output satisfies the constraint system, and because the training loss is the original objective function, the model can be trained self-supervised, without solved labels. On economic dispatch with perturbed loads, the end-to-end primal proxy keeps optimality gaps mostly below 1% on grids up to 30,000 buses while evaluating a batch of 256 instances in roughly 10 ms. On DC optimal power flow, the dual proxy produces geometric-mean dual gaps around 0.25-0.5%. On security-constrained optimal power flow with N-1 generator and line contingencies, the primal-dual proxy returns near-optimal dispatches in about 10 ms per instance, four orders of magnitude faster than the commercial solver baseline on systems with up to 6,515 buses.","pith_inferences":["A natural safety protocol the paper does not describe is to run the primal and dual proxies together and send any instance whose primal-dual gap exceeds a threshold back to an exact solver; most instances would still be served in milliseconds, and the exceptions would keep a formal guarantee.","The by-design feasibility guarantee is only as strong as the repair layer's coverage under distribution shift; a concrete stress test is to evaluate trained proxies on load patterns outside the training distribution and measure how the primal-dual gap degrades.","If automated synthesis of differentiable repair layers for general constraint classes were built, this approach would become a compiler-like technology; the paper itself lists the derivation of effective repair layers for a wide range of applications as an open issue."],"forward_implications":["Real-time risk assessment becomes practical: evaluating a 24-hour Monte-Carlo scenario of five-minute dispatches drops from roughly 15 minutes of solver time to about 5 seconds of proxy evaluation.","Self-supervised training removes the label bottleneck: for the largest tested security-constrained OPF case the reported training time of under three hours on a GPU is less than the time a solver needs to produce optimal labels for a single hard instance.","A dual proxy gives a cheap quality certificate: on DC-OPF the geometric-mean dual gap is around 0.25-0.5%, so a per-instance lower bound can be produced in milliseconds, not minutes.","The architecture pattern, raw prediction plus differentiable repair/completion, transfers across economic dispatch, linear programs, conic relaxations, and security-constrained OPF, which is why the paper presents it as a general methodology rather than a single model."],"supporting_citations":[{"why":"Supplies the E2ELR architecture and the economic-dispatch experiments behind the primal-proxy results.","marker":"[9]"},{"why":"Introduces the dual optimization proxy for linear programming and reports the DCOPF dual-gap results.","marker":"[23]"},{"why":"Introduces self-supervised primal-dual learning, the foundation of the PDL method.","marker":"[36]"},{"why":"Develops PDL-SCOPF and reports the contingency-dispatch experiments and speedups.","marker":"[37]"},{"why":"Provides the earlier deep-learning-plus-repair approach for SCOPF that the end-to-end framework extends.","marker":"[44]"},{"why":"Supplies the column-and-constraint-generation solver used as ground truth for SCOPF optimality gaps.","marker":"[45]"},{"why":"Provides the real-time risk-assessment application and the proxy-based Monte-Carlo evaluation.","marker":"[11]"},{"why":"Provides the benchmark power-system instances used to generate training and test data.","marker":"[6]"}],"fun_headline_variants":["Neural proxies make power-grid optimization 10,000x faster, always feasible","Feasible by design: learned proxies solve power grids 10,000x faster","Self-supervised optimization proxies: millisecond solutions, certified gaps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Every problem class of interest must admit a repair or completion layer that is both fast and differentiable; the paper gives working examples for dispatch, linear programs, and conic cases, but no general recipe that this layer always exists.","fun_headline_variants_meta":{"raw":{"variants":["Neural proxies make power-grid optimization 10,000x faster, always feasible","Feasible by design: learned proxies solve power grids 10,000x faster","Self-supervised optimization proxies: millisecond solutions, certified gaps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000613,"raw_usage":{"total_tokens":2813,"prompt_tokens":874,"completion_tokens":1939,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":1874}},"tokens_in":490,"tokens_out":1939,"duration_ms":16283,"temperature":1.0,"reasoning_tokens":1874,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:52:47.945924+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a random batch of held-out instances from the same load distribution, evaluate the trained primal and dual proxies on the same grid; the central claim fails if any primal output violates a hard constraint, or if the primal objective falls below the dual objective, since that would mean the dual proxy is not a valid lower bound.","supporting_citations":[{"cited_title":"End-to-end feasible optimization proxies for large-scale economic dispatch","cited_arxiv_id":null,"evidence_quote":"Supplies the E2ELR architecture and the economic-dispatch experiments behind the primal-proxy results."},{"cited_title":"Dual interior-point optimiza- tion learning, 2024","cited_arxiv_id":null,"evidence_quote":"Introduces the dual optimization proxy for linear programming and reports the DCOPF dual-gap results."},{"cited_title":"Self-supervised learning for large-scale preventive security constrained dc optimal power flow","cited_arxiv_id":null,"evidence_quote":"Develops PDL-SCOPF and reports the contingency-dispatch experiments and speedups."},{"cited_title":"Combining deep learning and optimization for preventive security-constrained DC optimal power flow","cited_arxiv_id":null,"evidence_quote":"Provides the earlier deep-learning-plus-repair approach for SCOPF that the end-to-end framework extends."},{"cited_title":"An exact and scalable problem decomposition for security-constrained optimal power flow","cited_arxiv_id":null,"evidence_quote":"Supplies the column-and-constraint-generation solver used as ground truth for SCOPF optimality gaps."},{"cited_title":"Real-time risk analysis with opti- mization proxies","cited_arxiv_id":null,"evidence_quote":"Provides the real-time risk-assessment application and the proxy-based Monte-Carlo evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the benchmark power-system instances used to generate training and test data."}],"review_version":1}