acceptodds
Under review as a conference paper at ICLR 2027

PRISM: From Feasible Geometry to Routing Capacity in Learning to Optimize

Abstract

Many learning-to-optimize methods struggle once the constraint set is curved, non-convex or disconnected: they either cannot guarantee hard feasibility there or lose solution quality. Feasibility alone is also not enough: a parametrisation can be feasible everywhere yet unable to represent the solution it is trained toward, and then no training closes the gap. For ray parametrisations, which place the output on a ray from an interior anchor, reachability is set by two budgets, anchors A and feasible segments K per ray: with one anchor and one segment, a second component or the far side of a hole is out of reach; each set admits a region of sufficient (A,K) budgets, and some are provably insufficient. PRISM realises such budgets: one shared network proposes ray candidates from an anchor bank, a parameter-free oracle marks each ray's feasible segments, and with valid anchors and an exact or certified oracle every hard-routed output is feasible. A measurement pass reads a data-supported (A,K) configuration off the feasible training labels and certifies its held-out coverage, without claiming the smallest budget. Under stated regularity assumptions we prove compact-set approximation, single-anchor obstructions, and a coverage-to-gap existence bound. On ten non-convex geometries the measured budgets cover at least 98% of held-out targets while single-anchor controls are starved on every disconnected set; on fifteen parametric problems PRISM meets each row's feasibility criterion and is best on twelve (the three legacy AC optimal power flow rows are coordinate-surrogate, box-only-residual evidence, not full-AC feasibility); on case30, conditional on a certified initial root, each accepted update certifies a continuous, strictly feasible root path with exact cost decrease.

Primary area:
optimization
Keywords:
Learning to Optimize, Constrained Optimization, Nonconvex Optimization, Hard-Constrained Neural Networks

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_jZCqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies geometric limitations of ray-based learning-to-optimize methods. It introduces ((A,K))-star-coverability, where (A) counts interior anchors and (K) counts feasible segments along each ray. PRISM uses a shared network and a ray-intersection oracle to generate feasible candidates, then selects one through hard routing. The paper provides conditional approximation and reachability results, a data-driven configuration procedure, and experiments on synthetic geometries and parametric optimization problems, including a separate certified AC-OPF extension.

Critical strengths
  • The distinction between feasibility, reachability, and trainability is useful. The geometric examples clearly demonstrate why training cannot overcome an insufficient representation.

  • The anchor–segment formulation is interpretable, and the controlled synthetic experiments support the benefits of increasing geometric coverage.

  • The authors explicitly disclose limitations, experimental selection procedures, and negative results.

Critical weaknesses
  • The presentation obscures the central contribution. The generic ray method, the specialized SOCP active-manifold configuration, and the case30 certified-path method have different mechanisms and guarantees. Their presentation under one framework makes it difficult to identify which results validate the proposed anchor–segment representation.

  • Some empirical results require independent confirmation. Appendix D.14 selects the better anchor descriptor retrospectively on the reporting split. Appendix D.13 also states that the SOCP-100D configuration was developed using data overlapping the reporting split. Furthermore, the three legacy AC-OPF rows evaluate coordinate-surrogate objectives and box residuals, rather than generation cost and full AC feasibility. These distinctions limit the interpretation of the aggregate “12/15” result.

  • Practical benefits remain conditional. PRISM requires valid interior anchors and an exact or certified intersection oracle, which may be expensive to obtain for complex non-convex problems. The synthetic experiments establish the geometric mechanism, but broader application benefits are less convincing. On certified case30, initial roots are available for 54/64 loads, and the numerical-gradient direction outperforms both learned directions in mean cost gap. This supports the certification mechanism more strongly than the benefit of learning.

Questions for authors
  • Can the authors provide results with descriptor and configuration choices fixed independently of the reporting data, separating the generic ray method from specialized extensions?

  • In the intended application setting, what is the end-to-end cost, including anchor acquisition and oracle evaluation, and what measurable benefit does learning provide?

3 comments

Thank you. We added tests with choices frozen before generating new inputs and report both descriptors. The revision distinguishes complete-predictor performance, learning attribution and costs within PRISM's feasibility-first design.

Weakness 1 - Distinct mechanisms. Generic PRISM uses valid anchors, ordered ray segments and hard selection. Specialized SOCP uses a numerical active-manifold proposal followed by ray readout, not path certification. Case30 certifies local feasible root paths and cost-decreasing re-anchoring. These are native extensions with separate contracts and evidence, not interchangeable theorems or latencies. The pooled "12/15" headline is removed.

Weakness 2 - Independent confirmation. Each task has 2,048 fresh instances. Saved checkpoints were frozen; missing slot checkpoints were reproduced with fixed recipes and identified. Both descriptors are reported. Coordinate-minus-slot losses are -0.380 [-0.527,-0.243] on Rings-8 and +0.039 [0.026,0.051] on Spiral; Disks-16 is unresolved (pointwise paired 95% intervals).

PRISM/HoP objective-minus-reference losses are 0.549/0.755 on TwoDisks, 0.122/0.330 on Annulus and 0.862/1.516 on Sin-QP; both pass 10−510^{-5}. These compare configured predictors, not one changed resource. Generic SOCP, specialized SOCP and OptNet give 16.16, 3.46 and 3.15. Legacy AC rows retain their surrogate-objective/box-residual scope. Fresh tests evaluate fixed choices without erasing historical selection exposure (Sec. 5.2, App. F.1).

Weakness 3 - Practical conditions. The ray-based method targets repeated solves of families with valid anchors and closed-form or soundly isolated one-dimensional crossings. We do not assume arbitrary nonconvex constraints provide cheap anchors/oracles. Curved paths use a separate local certificate. Initial roots are certified on 54/64 loads; the other ten remain unavailable, not proven infeasible. MOVE-GRAD and MOVE-NN both use our proposed certification/re-anchoring mechanism. The gradient policy has a lower mean cost gap (1.43% versus 1.55%/1.67%); this limits the learning claim, not the validity of that construction (Sec. 5.3, App. B).

Question 1 - Frozen choices and separate results. Yes: pipeline choices precede the new inputs, both descriptors are retained without test-set selection, and frozen checkpoints are distinguished from fixed-recipe reproductions. Generic rays, specialized SOCP and case30 have separate results (Weaknesses 1-2; App. F.1).

Question 2 - Cost and learning benefit. Learning: A separate fresh stream shares anchors, oracle, radial search and selection and counts every ray call. Learned directions beat random on five tasks, including four-times-budget random. At equal queries they beat both numerical proposers on Spiral (-0.052/-0.045) and Sin-QP (-2.64/-1.43). Numerical search is stronger on Disks-16, Rings-8 is unresolved, and NUMERIC is stronger on QCQP (+1.55). Larger budgets remove or reverse gains. These isolate finite-budget direction quality, not native radius/router performance (App. F.3).

Batch cost: For the first frozen checkpoint on existing E1 inputs, batch-512 five-pass median transfer-inference-return-check times on an A100 are 17.48/42.62 microseconds per input for QCQP/generic SOCP, versus 90.61/169.62 on one pinned CPU logical core. These are amortized batch times, not request latencies. Small batches and Sin-QP show no GPU benefit; HoP is faster on the matched grid. All five-pass ranges on the shared host are reported. Collation, loading, fixed-anchor acquisition and offline costs are excluded: this is a same-method hardware profile, not complete amortized cost or an equal-hardware solver speedup (App. F.8).

Including acquisition: Our historical case30 numerical service and sequential local-reference service return checked states on the same 57/64 loads. Three-repeat mean times for all 64 attempts are 2.204/28.795 s, including 1.607 s for our acquisition; the 2.214% mean cost premium concerns 57 common successes. This exposed engineering workload is not fresh confirmation, a quality-matched speedup or certified-path timing (App. B.3).

Components: On 225 stored selection states, network-only proposals average 0.18 ms versus 1.19 ms for gradients. The current NN arm computes both; prefix certification dominates separately recorded trajectory costs. No full no-gradient certified trajectory was tested. Serial speculative acquisition reduces solver calls without a consistent time gain (App. D.15, F.8).

Thanks for your efforts. I will maintain my assessment.

Thank you for taking the time to read our response. We appreciate your feedback on separating the generic ray method from the specialized extensions and reporting the evaluation with choices fixed before testing.

Reviewer_mPA6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PRISM, an always-feasible (theoretically) framework for learning to optimize. The authors introduce (A,K)(A,K)-star-coverability to characterize the geometric reachability of ray-based representations, where AA denotes the number of anchors and KK the number of feasible intervals per ray. PRISM combines a shared neural network, a ray-intersection oracle, and hard routing to generate feasible solutions. The paper establishes approximation guarantees under regularity assumptions, proves geometric reachability lower bounds, and develops a data-driven procedure for selecting anchor–segment configurations. Extensions to OPF seem new but are somehow irrelated with the main content.

Critical strengths
  • Interesting geometric perspective: The paper distinguishes feasibility from geometric reachability and trainability. The (A,K)(A,K)-star-coverability framework provides an intuitive way to characterize structural limitations of ray-based representations. It's a necessary and somehow sufficient framework to model feasibility in ray-based settings.

  • Meaningful theoretical results: The paper establishes impossibility results for insufficient routing budgets and connects geometric coverage to approximation guarantees and expected optimality gaps under explicit assumptions.

Critical weaknesses
  • Limited empirical validation beyond method-aligned synthetic benchmarks: A substantial portion of the experimental suite consists of constructed geometries, such as unions of disks, nested annuli, and spirals, whose structural properties directly motivate the proposed anchor–segment representation. While these controlled experiments are appropriate for validating the geometric theory, they provide limited evidence of the method's broader practical advantages. Most non-AC parametric benchmarks are also synthetic, while the three legacy AC-OPF comparisons use surrogate objectives and do not establish full AC feasibility. Moreover, the certified AC-OPF experiments do not demonstrate a clear advantage of learned directions over a numerical gradient baseline. Evaluations on established, independently motivated optimization benchmarks with realistic constraints and original objectives would substantially strengthen the empirical contribution.

  • An incorrect regularity claim in Definition 3.1: The authors claim that the intersection of a ray with a feasible set having a piecewise-smooth boundary is a finite union of closed intervals. This is false, even for compact sets with C∞C^\infty boundaries. Smooth oscillatory boundaries can produce infinitely many disjoint ray intervals, potentially accumulating at a finite point. In such cases, the intervals may not even admit the increasing enumeration required to define sa(y)s_a(y). The authors should explicitly assume ray-finiteness or restrict the geometric class, for example to semialgebraic sets.

  • Insufficient justification of regular-cover conditions: At the end of the ray-cover budget region discussion, the authors suggest that their closed-form oracles satisfy the required regularity whenever the constraint data depends continuously on xx. However, continuity of constraint data alone does not guarantee continuous anchor selections, locally constant interval counts, or jointly continuous interval endpoints. A simple case is considering y2≤xy^2 \leq x and x=0x=0.

Questions for authors
  • To what extent do the geometric obstructions demonstrated on the constructed benchmarks arise in realistic constrained optimization problems? Can the authors evaluate PRISM on established benchmarks not specifically designed around disconnected components or multiple ray crossings, using their original objectives and constraints?

  • Can the authors provide stronger evidence that the learned components of PRISM improve solution quality on realistic high-dimensional tasks, independently of the feasibility oracle and geometry-specific numerical procedures?

3 comments

Thank you. Both mathematical objections are correct and have been repaired. PRISM remains a feasibility-first L2O framework, with geometric access, trained-predictor quality and nonlinear-equality extensions evaluated separately.

Weakness 1 - Empirical scope. Our suite separates geometric diagnostics, parametric optimization families and an existing physical application. Disks, rings and spiral diagnose geometry; TwoDisks and Annulus also involve a disconnected set and a hole. QP-eq uses a DC3-style formulation; the convex QCQP/SOCP generators document HomeoProj-inspired adaptations. These are generated instances, not unchanged public benchmark collections (App. D.1).

After freezing choices, we evaluated 2,048 fresh instances/task. PRISM/HoP objective-minus-reference losses are 0.549/0.755 on TwoDisks, 0.122/0.330 on Annulus and 0.862/1.516 on Sin-QP. Convex QCQP/SOCP give 34.61/34.97 and 16.16/16.23; one valid anchor and one segment suffice geometrically there. Both methods pass 10−510^{-5} on these rows. These are configured-predictor comparisons, not matched-compute or single-factor ablations. Specialized SOCP is reported separately (App. F.1).

The existing case30 study uses original generation cost and full AC constraints. Certified Path verifies a continuous feasible prefix; Re-anchoring accepts a point on it only after root-identity and exact cost-decrease checks. Rejection retains the current root. Certified initial roots were obtained for 54/64 loads; the other ten are unavailable, not proven infeasible. This establishes a curved-path construction under its rational-model contract (Sec. 5.3, App. B). The three legacy AC rows remain surrogate-objective/box-residual comparisons, not evidence of full AC feasibility or generation-cost performance (App. B.3).

Weakness 2 - Ray-finiteness. Definition 3.1 no longer infers finite ray components from smooth boundaries or assumes an increasing enumeration. A target's index counts components meeting the anchor-to-target prefix, using ∞\infty for every infinite count. Finiteness is assumed where the approximation proof requires it; semialgebraicity supplies a sufficient condition for the relevant benchmark sets. Singleton and unbounded components are treated explicitly (Def. 3.1, App. A.1).

Weakness 3 - Regular covers. We removed the implication from continuous constraint data. The revised assumptions require a fixed affine direction space, continuous valid anchor maps, and finite locally constant interval counts with continuous nondegenerate endpoints outside closed exceptional sets with empty interior. The simple-crossings lemma supplies a local sufficient test, including fixed Boolean combinations such as disk/annulus unions. Per-instance anchor validity does not imply the continuous selector needed for uniform approximation. Approximation requires a complete, ordered exact oracle; sound inner intervals preserve candidate feasibility but may omit targets. The corrected hypotheses retain the structural obstructions, conditional approximation and coverage-to-gap results you identified as strengths (Sec. 4, App. A).

Question 1 - Established benchmarks. This revision does not add evaluation on an additional established public benchmark with its original objective and constraints. The controlled geometries test structural predictions, not their prevalence in applications. Existing case30 uses the original objective and full constraints, but tests certified curved paths rather than generic ray obstructions (Sec. 5.3, App. B).

Question 2 - Independent learning benefit. Common-layer tests fix anchors, oracle, radial search and selection and count every ray query. Learned directions beat random on five tasks, including QCQP-100D and four-times-budget random search. They beat both numerical proposers at equal queries on Spiral and Sin-QP; numerical search is stronger on Disks-16 and QCQP, Rings-8 is unresolved, and larger budgets reverse some gains (App. F.3).

MOVE-GRAD uses the same proposed certification/re-anchoring framework as MOVE-NN, not an external replacement. Its lower mean cost gap (1.43% versus 1.55%/1.67%) supports that numerical-policy instance, not a learned-policy advantage. Thus we establish specific finite-budget proposal benefits, not realistic high-dimensional superiority; the curved-path contribution is assessed on its own guarantees and evidence.

Reviewer_mPA6OpenReview ↗

Response to the rebuttal

Thanks for your response. I would like to maintain my assessment as only part of the concerns are addressed.

Thank you for taking the time to consider our response. If there is a particular concern you feel the revision has not addressed, we would appreciate knowing which one.

Reviewer_QfM9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the challenges faced by existing learning to optimize methods in simultaneously ensuring feasibility and optimal solution reachability in nonconvex and disconnected feasible regions. To address these challenges, it proposes PRISM, a framework that expands the reachable solution space through multiple anchors, multiple feasible ray segments, and a feasibility oracle, while employing a data driven approach to select parameter configurations. Theoretically, the paper analyzes the approximation capability, reachability limitations, and the relationship between coverage and optimality gaps. Experiments on various nonconvex geometries and parametric optimization problems demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The research problem is well motivated. The paper not only focuses on whether neural networks can generate feasible solutions but also investigates whether the underlying parameterization can reach high-quality solutions, addressing a practically meaningful problem.
  2. The theoretical analysis is relatively comprehensive. The paper introduces (A,K)-star-coverability and establishes theoretical connections among approximation capability, structural reachability limitations, coverage, and optimality gaps.
  3. The proposed method has broad applicability. By combining neural networks with a feasibility oracle, the framework guarantees feasible outputs under certain conditions and can handle various nonconvex constrained optimization problems.
Critical weaknesses
  1. The computational cost and scalability of the oracle remain unclear. For complex high-dimensional nonconvex constraints, accurately computing multiple intersections between a ray and the feasible region can be challenging, potentially limiting the practical applicability of the proposed method.
  2. There is a gap between theoretical guarantees and practical optimization performance. The theoretical results mainly establish the reachability of the parameterization but do not guarantee that neural networks can learn to find high-quality solutions. Even when target solutions are reachable, training difficulties, incorrect candidate selection, or insufficient network optimization may prevent the model from finding high-quality solutions.
  3. The stability of the anchor selection strategy is insufficiently demonstrated. Experiments show that different anchor configurations can lead to substantial differences in solution quality, even when their coverage rates are similar. Moreover, some random configurations outperform the proposed data-driven selection strategy.
  4. The independent contribution of the learning component remains unclear. In some experiments, the directions predicted by neural networks do not outperform numerical gradient directions. It remains unclear whether the performance improvements primarily come from the learned optimization strategy or the geometric parameterization.
Questions for authors
  1. The paper mainly establishes existence guarantees relating reachability to optimality gaps, but these results do not directly guarantee the performance of trained models. Could the relationship between reachability and trainability be further analyzed?
  2. Experiments show that different anchor configurations can achieve similar coverage rates but substantially different solution quality. How can the stability of the data driven anchor selection strategy and its generalization across different data distributions be ensured?
  3. The performance improvements of PRISM may arise from the multi anchor parameterization, the feasibility oracle, or the neural learning component. Could more comprehensive ablation studies and comparisons with traditional optimization methods be conducted to clarify the individual contributions of these components?
  4. PRISM relies on an exact or certified Ray Oracle, while computing feasible intervals in high dimensional nonconvex problems can be computationally expensive. Could the scalability and effectiveness of the proposed method be further investigated when an exact oracle is unavailable and only approximate feasibility information can be obtained?
  5. The proposed framework involves several design choices and hyperparameters, including the number of anchors, feasible segments, ray directions, and anchor feature representations. Could the sensitivity to these configurations be further analyzed, and could more systematic guidelines for parameter selection be provided to reduce reliance on empirical tuning?
3 comments

Thank you. PRISM targets high-quality feasible predictions for repeated constrained problems. The revision separates geometric access, candidate quality and selection, and tests each contribution.

Weakness 1 - Oracle cost. Fixing an anchor and direction reduces intersection search to one dimension. Our linear, quadratic and conic instances have closed-form crossings; other constraints may require root isolation. One dimension alone does not ensure cheap enumeration or certified Newton roots.

At batch 512, five-pass median transfer-inference-return-check times for the first frozen checkpoint are 17.48/42.62 microseconds per input on an A100 for QCQP/generic SOCP, versus 90.61/169.62 on one pinned CPU logical core. These are same-pipeline, batch-amortized hardware measurements on a shared host, not request latencies or algorithmic speedups. Small batches and Sin-QP show no GPU advantage. Acquisition, collation, loading and offline costs are excluded; full ranges are reported (App. F.8).

Weakness 2 - Trained performance. Reachability is not a training guarantee; it identifies restrictions training cannot remove. After freezing choices, 2,048 fresh instances/task gave PRISM/HoP objective-minus-reference losses of 0.549/0.755 on TwoDisks, 0.122/0.330 on Annulus and 0.862/1.516 on Sin-QP. Both pass 10−510^{-5} on these rows. These compare configured predictors, not single-factor ablations. Changing only selection among fixed candidates reduces Rings-8 loss from 0.769 to 0.201; QCQP has no observed selection regret (App. F.1-F.2).

Weakness 3 - Anchor variability. Construction prioritizes valid interior anchors and visibility, with a short training probe among similarly covered configurations. It does not optimize over all possible banks. In separate matched retraining, fixing size, segment budget, width and recipe within each construction pool, our procedure improves average loss over random banks by 0.72 on Disks-16 and 0.28 on Spiral. Spiral's gain concentrates in one pool. Historical random-bank replacements used frozen measured-bank networks, not per-bank retraining; the two experiments are distinguished (App. F.4).

Weakness 4 - Learning contribution. A second fresh stream fixes anchors, oracle, radial search, selection and budgets, counting every ray query. Learned directions beat random on five tasks, including random at four times the queries. At equal queries, learned-minus-comparator differences against the numerical trajectory/two-start search are -0.052/-0.045 on Spiral and -2.64/-1.43 on Sin-QP, with pointwise paired 95% intervals excluding zero. Numerical search wins on Disks-16, Rings-8 is unresolved, and NUMERIC wins on QCQP. Larger numerical budgets remove or reverse some gains: the benefit is task-specific and finite-budget (Sec. 5.2, App. F.3).

Question 1 - Reachability versus trainability. We separate geometric restriction, proposal quality and selection. The fixed-candidate diagnostic above isolates selection; approximation remains an existence result, not training convergence (App. A.3, F.2).

Question 2 - Stability and distribution shift. Matched retraining supports an average benefit within tested pools, not uniformly good losses for all covered banks. Distribution-shift robustness remains unestablished (App. F.4).

Question 3 - Ablations. Yes: common-layer direction comparisons, fixed-candidate selection tests and matched bank retraining isolate different effects. PRISM/HoP pipeline comparisons are not treated as single-factor A/K/MA/K/M ablations (App. F.1-F.4).

Question 4 - Approximate oracles. With valid anchors and hard selection, sound inner intervals preserve candidate feasibility in real arithmetic but may lose coverage. Approximation needs complete, ordered exact intervals and regularity. Some checked static outputs fail exact membership while passing 10−510^{-5}; unvalidated interval variants remain certification-pending. Our timings do not establish cheap certification for arbitrary constraints (App. F.6, F.8).

Question 5 - Configuration. Use development visibility for A,KA,K, a short probe for similarly covered banks, development quality/cost for MM and descriptor, then freeze before testing. Independent sound verification gives nine-bank joint 95% coverage lower bounds above 0.93. Coordinates win on Rings-8 and slots on Spiral; neither descriptor wins everywhere (Sec. 3.3, App. F.1, F.4, F.7).

Thank you for your detailed response and the additional experiments. I appreciate the substantive improvements made in addressing my concerns. However, I still have reservations regarding the scalability of the proposed framework and the gap between theoretical guarantees and practical optimization performance. Therefore, I have decided to maintain my original rating.

Thank you for taking the time to review the new experiments. We appreciate your acknowledgment of the improvements and would like to clarify what the revision establishes on the two remaining points.

Oracle scalability. Fixing an anchor and a direction reduces the intersection problem to one dimension. The tested linear, quadratic, and conic constraints admit closed-form crossings, but general nonlinear constraints can still make certified interval enumeration expensive. The batch profiles in Appendix F.8 characterize the tested implementations for repeated solves; they do not establish low oracle cost for arbitrary constraints.

From reachability to trained performance. The theory identifies which solutions the parameterization can reach. We test separately whether trained models find better solutions: after freezing all pipeline choices, PRISM achieved lower objective-minus-reference loss than our HoP reimplementation on TwoDisks, Annulus, and Sin-QP, each evaluated on 2,048 fresh instances. In the separate direction ablation, we fixed the anchors, oracle, radial search, and selection rule and counted every ray query. Learned directions beat both numerical search baselines at equal query budgets on Spiral and Sin-QP, though not across all tasks. This isolates a benefit of learning in those settings without assuming that reachability ensures successful training (Section 5.2; Appendices F.1–F.3).

If there is a particular constraint family or comparison that you think would most directly test either concern, we would appreciate your perspective.

Comments on the paper

1 comment

We thank the reviewers for their careful reading. PRISM is a feasibility-first learning-to-optimize framework for repeated constrained problems. Multiple anchors and ordered feasible ray segments expand geometric access; multiple directions provide candidate solutions. The revision clarifies how each mechanism, guarantee and experiment supports this design.

Independent evidence for trained performance and learning. With pipeline choices frozen before generating 2,0482,048 fresh instances per task, PRISM achieves lower objective-minus-reference loss than our HoP reimplementation on TwoDisks (0.5490.549 vs. 0.7550.755), Annulus (0.1220.122 vs. 0.3300.330), and Sin-QP (0.8620.862 vs. 1.5161.516). These compare configured predictors, not single-factor ablations. Separate common-layer tests hold the geometry, oracle, radial search and selection fixed and count every ray query: learned directions outperform random across five tasks and both specified numerical proposers on Spiral and Sin-QP at equal query budgets. The gains are task- and budget-dependent. Both descriptors are reported, and the pooled 12/1512/15 result is no longer a performance headline (Sec. 5.2; App. F.1–F.3).

Corrected assumptions, retained theoretical results. We corrected the ray-finiteness assertion and made regular-cover assumptions explicit. The structural reachability, conditional approximation and coverage-to-gap results are retained under their stated assumptions. Complete, ordered exact oracles and sound inner oracles have different guarantees; mathematical feasibility is distinguished from numerical output checks, and reachability is not a training-convergence guarantee (Sec. 4; App. A).

Nonlinear equalities and measured batch performance. The existing case30 AC-OPF study certifies continuous feasible paths and cost-decreasing re-anchoring under its rational-model contract, conditional on certified initial roots (54/6454/64 loads). Both learned and numerical direction policies use this construction; the result supports certified feasible exploration, not a learned-policy advantage (Sec. 5.3; App. B).

New batch profiles include transfers, inference, output return and numerical checks. At batch 512512, the tested QCQP and generic-SOCP pipelines have higher throughput on an A100 than on one CPU logical core; small batches and Sin-QP do not benefit. These are same-pipeline hardware comparisons, not algorithmic speedups; anchor acquisition and offline costs are excluded (App. F.8).

Our reviewer-specific responses address each Weakness and Question in its original order and identify the corresponding manuscript changes.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.