acceptodds
Under review as a conference paper at ICLR 2027

Convex-SKM-Net: Low-Latency Block Halfspace Corrections for Neural Convex Constraint Satisfaction

Abstract

Fast inference and strict constraint satisfaction are essential for learning-based decision-making. Existing methods often rely on constraint-specific projections or feasible-set parameterizations, while iterative repair can be costly. We introduce Convex-SKM-Net, a convex constraint-satisfaction layer that forms linearization halfspaces from residual and (sub)gradient oracles without constraint-specific inequality projectors. It applies fixed-depth dual projected-gradient descent (dual PGD) to minimum-displacement block problems using GPU-friendly matrix operations. We jointly learn the predictor and a positive-definite mobility, adapting the correction geometry to improve objective quality. We prove convergence of the outer iterates at fixed dual depth and linear convergence in distance under Slater's condition and bounded subgradients. All our test outputs on three second-order-cone model-predictive control (SOC-MPC) horizons and six power-system convex relaxations satisfy a feasibility tolerance. Comparisons with the latest constraint-satisfaction methods show inference speedups of up to on SOC-MPC and on a convex quadratically constrained quadratic programming (QCQP) benchmark with joint training. On QCQP, the mean objective gap is %, with full test feasibility at . Learning mobility reduces the five-seed affine-task mean gap by % and improves gaps in matched power-system comparisons.

Primary area:
optimization
Keywords:
Convex constraint satisfaction, constrained neural networks, Sampling Kaczmarz–Motzkin, block halfspace corrections corrections, dual projected gradient, GPU acceleration

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_NHX2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Convex-SKM-Net is a differentiable feasibility-repair layer for affine equality and convex inequality constraints. It enforces equalities, constructs linearization halfspaces, and approximately solves minimum-displacement block subproblems using fixed-depth dual projected-gradient updates. A learned positive-definite mobility adapts the correction geometry. The paper establishes convergence under suitable assumptions and evaluates the method on control, power-system convex relaxations, and optimization benchmarks.

Critical strengths
  • Practical generality: The method requires constraint residuals and subgradients rather than constraint-specific inequality projectors. Block updates and reused linearizations support efficient GPU execution.

  • Relevant theoretical analysis: The finite-inner-depth progress bounds address an important tradeoff between correction accuracy and inference cost.

  • Substantial empirical evaluation: Experiments report objective quality, original-constraint feasibility, and latency, with useful baseline sweeps, matched-prediction comparisons, and mobility ablations.

Critical weaknesses
  • The incremental novelty needs clearer justification. Differentiable prediction-and-repair frameworks already exist in DC3, while halfspace approximations and block subgradient projections are established convex-feasibility techniques. GLinSAT also provides a differentiable, matrix-factorization-free constraint layer using dual first-order optimization. The paper should more precisely identify what is new beyond extending these ideas to nonlinear convex constraints and learning the correction geometry.

  • In particular, one zero-initialized dual step gives Δ=−αU⊤[s]+\Delta=-\alpha U^\top[s]_+, which equals one gradient step on the local squared-violation objective. This does not make the complete methods equivalent, but motivates isolating the benefits of multiple dual steps and the other design choices.

  • Important inference configurations fall outside the stated guarantees. Appendix B acknowledges that the one-step power estimates used in ACOPF and QCQP inference may violate the theoretical step-size condition. A matched comparison with a certified safe rule is needed to establish the performance cost of retaining the convergence guarantees.

  • The source of the performance gains is not fully isolated. In matched synthetic ablations, dual PGD is only approximately 1.06–1.17 times faster than Violation-GD. Additional comparisons of correction distance, downstream objective quality, and robustness would clarify its specific advantage. GLinSAT is cited but not evaluated; comparison on a bounded, linearly constrained benchmark would be useful.

  • Some empirical claims need stronger support. Traditional solvers and neural methods use different hardware, and SOC-MPC solver evaluations do not use warm starts. QCQP uses only one training seed; repeated timing passes do not establish training robustness.

Questions for authors
  • How do latency, feasibility, and objective quality change with theoretically safe inference step sizes?

  • What specifically distinguishes the finite-depth analysis from existing inexact block-projection results?

  • Can the authors add GLinSAT and matched violation-gradient repair with learned preconditioning as baselines?

1 comment
AuthorsOpenReview ↗

Responses to Reviewer NHX2

Thank you for the detailed review and for noting the generality of the oracle interface and the relevance of the finite-depth bounds.

Safe inference steps (Q1). We held each checkpoint fixed and replaced power1 with ω/ℓB\omega/\ell_B, where ℓB\ell_B bounds the maximum absolute Gram row sum for the mobility-transformed, normalized rows. We used ω=1.8\omega=1.8 on QCQP and 1.95 on ACOPF. Both rules met tolerance on every input in these comparisons. For dual PGD on QCQP (seed 456, 10−810^{-8}), power1 averaged 3.026 ms (0.01155% gap) and Gershgorin 1.8 averaged 3.826 ms (0.01213%), excluding transfers. On SOCP300, both step rules met 10−310^{-3} on all 1,000 inputs for each seed. We clipped negative per-input gaps to zero before averaging.

Seed (power1 → safe)Mean msP99 msMean gap
129.9 → 82.4447 → 8890.0388% → 0.0405%
224.4 → 73.9270 → 8350.0343% → 0.0366%

The safe rule was slightly faster on the 57-bus tasks. We will tabulate mean and P99 latency for all tasks.

Finite-depth analysis (Q2). Our bounds apply after a fixed TT even if a block inequality remains violated. We retain T-SKM-Net's equality projection. Residuals and subgradients define a block of halfspaces, and we take TT dual PGD steps on its minimum-displacement QP. Kiwiel (1995) aggregates inequalities into a surrogate halfspace. Necoara and Nedić (2021) analyze minibatch subgradient corrections. DC3 takes gradient steps on constraint violation, and GLinSAT uses an entropy-regularized formulation for bounded linear constraints. Díaz Millán et al. (2021) stop an inner solver at a feasible approximate projection using an error test.

With zero initialization and step size 1/LB1/L_B, Thm. 4.1 guarantees a decrease of at least cB,T>0c_{B,T}>0 in squared distance to every feasible point whenever the block is violated, for any T≥1T\ge1. The first top-bb block captures at least b/mb/m of the total squared normalized violation. Under the assumptions of Cor. 4.2 this gives outer convergence, with a linear distance rate under Slater's condition. In the analysis coordinates, Eq. (24) bounds the squared error relative to the fixed block's exact displacement Δ⋆\Delta^\star by χB∥Δ⋆∥2/(T+χB)\chi_B\lVert\Delta^\star\rVert^2/(T+\chi_B), where χB\chi_B depends on that block. As you note, with the same rows and step size the first dual step from zero equals a squared-violation gradient step. At T=1T=1 the measured mean gaps agree to roundoff.

Matched VGD and GLinSAT comparisons (Q3). On QCQP (seed 456), we jointly trained one model per update and used its training update at test time. Both met 10−810^{-8} on all 833 inputs. In the same RTX 4090 run as Q1, dual PGD (power1) averaged 3.026 ms (0.01155% gap) and VGD (power1) 3.448 ms (0.01543% gap). When trained separately on the affine task, dual PGD and learned-preconditioned VGD gave five-seed mean gaps of 0.159% and 0.172%, respectively.

In 96 QCQP post-processing comparisons (four predictors, four tolerances, two step rules, T = 5, 10, or 20), both repairs passed all 833 inputs. Relative to VGD, dual PGD reduced mean latency by 3.37–11.60% and mean displacement by 0.0011–0.0096%. Mean signed gap fell by 0.000686–0.003474 percentage points. From matched starts in nine affine stress settings, dual PGD reached 10−310^{-3} within 100 outer iterations on 461/576 cases (80.0%) and VGD on 384/576 (66.7%).

On a bounded linear task (16 variables, five seeds), we used the authors' GLinSAT implementation. All three methods ran on the same TorchAX/XLA stack on RTX 4090s and met 10−310^{-3} on all 64 test inputs in each seed.

MethodMean abs. objective excessBatch-one ms
GLinSAT0.1029.53
Ours, identity0.0830.30
Ours, learned mobility0.1060.35

Solver timing, warm starts and seeds. Neural methods ran on RTX 4090 GPUs. SOC-MPC solvers used one CPU thread. ACOPF used MOSEK's default threading (App. D.5). The original SOC-MPC contexts were independent. At H=80H=80, a reference controller generated common trajectories for three initial deviations and two feasibility tolerances (10−3,10−610^{-3},10^{-6}). For COSMO, we set both eps_abs and eps_rel to 10−510^{-5} in the looser test and 10−810^{-8} in the tighter one. It reused its workspace and the previous primal and dual states. Timing excludes the first step of each trajectory and includes transfers for the GPU method. All methods met the external tolerance in all six settings. At 10−310^{-3} and the smallest deviation, warm starting reduced COSMO's mean latency from 25.5 to 4.00 ms, versus 4.48 ms for our model. Our model was faster in the other five settings. Its mean absolute relative objective error was 0.73–0.97%. We will report both methods' objective errors in the full table.

Four additional QCQP training seeds bring the total to five. Each passed all 833 inputs at 10−610^{-6}. The gap (mean ± sample SD across seeds) was 0.0122% ± 0.0014% with power1 and 0.0118% ± 0.0003% with the safe rule.

We hope these results address your concerns.

Reviewer_vZhNOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a neural repair layer for convex-constrained decisions. It constructs local supporting halfspaces, selects blocks of violated constraints, and applies fixed-depth dual PGD updates. A learned mobility matrix modifies correction directions. The paper provides convergence analysis under stated assumptions and reports empirical improvements in feasibility, objective quality, and inference latency across several convex optimization tasks.

Critical strengths

The method provides a unified repair procedure for different convex constraints without requiring a dedicated projection operator for each constraint type. The reported feasibility and latency results demonstrate potential practical value for neural decision systems.

Critical weaknesses
  1. The computational advantage is not sufficiently established beyond the evaluated instances. Fixed-depth block PGD does not eliminate iterative optimization: total cost still depends on constraint evaluation, block selection, block construction, and the number of outer corrections. These costs may offset the speed advantage of the neural predictor. The evaluation does not sufficiently establish behavior on narrow or poorly conditioned feasible regions requiring repeated block reselection. Consequently, the reported average speedups do not resolve the main concern about reliable low-latency inference.

  2. The guarantees are narrower than the practical benefits being sought. Convergence under the stated assumptions does not by itself establish feasibility within a small inference budget. Moreover, learned mobility provides empirical objective-quality improvements rather than an optimality or objective-degradation guarantee. The supporting-halfspace argument fundamentally relies on convexity, restricting applicability to convex problems or convex relaxations. Taken together, the contribution is a useful convex feasibility-repair mechanism, but its broader significance and predictable finite-budget performance remain insufficiently established.

Questions for authors
  1. Complexity clarification: What is the full inference complexity, including constraint and gradient evaluation, block selection, mobility operations, and all outer iterations? How does each component scale with problem dimension and constraint count?

  2. Practical convergence: What do the theoretical bounds imply for the number of corrections needed to reach the experimental feasibility tolerances? Are these bounds informative at the reported inference budgets?

  3. Difficult-instance evidence: How does the method behave on nearly infeasible or narrow feasible regions with strongly correlated constraints? Please report iteration counts, block reselections, failure rates, and tail latency, rather than only average runtime.

  4. Mobility guarantees: Beyond empirical average objective gaps, is there any control on objective degradation caused by the correction or learned mobility?

1 comment
AuthorsOpenReview ↗

Responses to Reviewer vZhN

Thank you for the careful review and for noting that the method gives one repair procedure for different convex constraints without a dedicated projector for each type.

Q1. Full inference cost. Total inference cost is the cost of prediction and equality projection, plus the sum over the outer iterations and a final residual check. In null-space coordinates, the cost per outer iteration is (App. B.2, Eq. 58) O(Clin+S[mq+Csel(m,b)+b2q+Tb2]).O\left(C_{\rm lin}+S[mq+C_{\rm sel}(m,b)+b^2q+Tb^2]\right). Here ClinC_{\rm lin} covers residual and subgradient evaluation, coordinate transforms and row normalization. For SDP this includes finding a minimum eigenpair for each PSD block. Each of the SS block updates scores mm rows in dimension qq, selects bb rows at cost Csel(m,b)C_{\rm sel}(m,b), forms their Gram matrix and takes TT dual steps. In our implementation, diagonal mobility scales the columns of the equality-projected gradients. We accumulate updates in ambient dimension nn, so the scoring and Gram terms become mnmn and b2nb^2n.

For the 57-, 118- and 300-bus SDP relaxations, mean end-to-end times were 4.93, 12.7 and 184 ms on the RTX 4090 (batch size one). MOSEK took 192, 358 and 991 ms on CPU with its default thread settings, excluding JuMP model construction (Table 3 and App. D.5).

Q2. Corrections needed to reach tolerance. The bounds do not certify the budgets used in our experiments, because we did not estimate their constants for individual test inputs. They give the number of outer iterations in terms of problem constants. Under Eq. (41), K≥M2d02/(w‾ε2)K\ge M^2d_0^2/(\underline w\varepsilon^2) outer iterations suffice for at least one iterate to reach tolerance ε\varepsilon on the original inequalities. Here MM bounds the subgradient norms, d0d_0 is the initial distance to feasibility in the analysis coordinates and w‾\underline w is the descent bound for the chosen step rule. By convexity, the maximum violation satisfies vK≤MdKv_K\le Md_K, where dKd_K is the distance to feasibility at iterate KK. Under Slater's condition, the distance rate in Cor. 4.2 then gives vK≤Md0(1−γ)K/2v_K\le Md_0(1-\gamma)^{K/2}, where γ=w‾/κ22\gamma=\underline w/\kappa_2^2 depends on the error-bound constant.

Test setToleranceStep ruleMax. outer iterations
QCQP post-processing, 4 seeds, 833 inputs per seed10−410^{-4}power12
SOCP118, seed 1, 400 inputs10−310^{-3}power158
Same SOCP118 checkpoint and inputs10−310^{-3}Gershgorin139

Q3. Narrow feasible regions and tail latency. We tested 100 affine constraints at nine combinations of row correlation and half-width, with 64 known-feasible inputs per setting. Both repairs started from the same predictions. At the paper's affine tolerance of 10−310^{-3} and a budget of 100 outer iterations, dual PGD missed tolerance on 115/576 inputs (19.97%) and Violation-GD (VGD) on 192/576 (33.33%). The table shows the three settings with any misses. Both methods passed all 64 inputs in each of the other six.

CorrelationHalf-widthPassed (PGD / VGD), each of 64PGD inputs using all 100 roundsMean PGD reselectionsPGD P99 ms (passed / failed)
0.9910−210^{-2}64 / 0090.20333.835 / —
0.9910−410^{-4}13 / 056491.75044.728 / 44.150
0.999910−410^{-4}0 / 064478.156— / 44.910

We count a reselection when the selected set changes between updates. Each P99 is computed across inputs after averaging three timings per input. In the middle row, 5 of the 56 inputs that used all 100 rounds met tolerance on the last round.

In QCQP post-processing, dual PGD reached 10−410^{-4} on all 833 inputs for each of four seeds (power1, T=10T=10). We averaged three timings per input before computing P99. Across seeds, P99 was 1.079–1.200 ms for dual PGD and 1.182–1.332 ms for VGD.

Q4. Objective change during repair. Yes. The bound depends on how far the raw prediction y^\hat y lies from feasibility. Under the theorem's step-size condition, if fxf_x is LfL_f-Lipschitz along the segment from y^\hat y to the repair iterate yy, then ∣fx(y)−fx(y^)∣≤2Lfcond(Pz)⋅dist2(y^,Y),|f_x(y)-f_x(\hat y)|\le 2L_f\sqrt{\mathrm{cond}(P_z)}\cdot\mathrm{dist}_2(\hat y,\mathcal Y), where Y\mathcal Y is the feasible set and cond(Pz)≤eℓ\sqrt{\mathrm{cond}(P_z)}\le e^\ell by the mobility parameterization (Eq. 53). For fixed mobility PzP_z, Fejér monotonicity bounds the null-space displacement, which is orthogonal to the equality-projection step. We will include the proof in the appendix. The bound applies to the SOC-MPC steps and the certified Gershgorin rule. For power1, our evidence on objective quality is empirical. This argument allows a nonconvex objective but assumes convex inequalities and affine equalities. In ACOPF it therefore applies to the convex relaxations.

We hope these results address your concerns.

Reviewer_DQL1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a convex constraint satisfaction layer named Convex-SKM-Net. The underlying idea is to construct linearization halfspaces and apply a fixed-depth dual PGD to minimum-displacement problems, with a key claimed advantage being avoiding constraint-specific inequality projections. The parts that are learned are the predictor and a positive definite matrix which controls the directional displacement. Experiments on quadrotor SOC-MPC, SOCP/SDP relaxations of ACOPF, an affine-constrained nonconvex task and a convex QCQP benchmark report full feasibility under the defined tolerances with substantially lower inference latency than recent constraint enforcement methods.

Critical strengths
  1. The method is technically sound. The convexity assumption ensures that the linearization halfspaces contain the feasible set, the block QP is always feasible and the dual PGD inner solver with a fixed depth keeps the updates GPU-friendly.
  2. In all benchmarks (SOC-MPC, ACOPF, an affine nonconvex task and a convex QCQP benchmark), the outputs meet the tolerances and inference is significantly faster than the Π\Pi-net, LMI-Net, and SnareNet baselines.
Critical weaknesses
  1. The first main concern is that the contribution is an assembly of well-established components (subgradient linearization halfspaces, block Kaczmarz–Motzkin corrections and a learned diagonal matrix) on top of the T-SKM-Net pipeline, and in addition, it is unclear whether the conceptual advantages claimed (generality through oracles, strict feasibility, and a provable benefit from block corrections) actually hold up under closer inspection.

    To begin with, relative to T-SKM-Net, the authors replace linear rows with subgradient linearization halfspaces, single-row random sampling with greedy top-b blocks solved by a few dual PGD steps and the Euclidean norm with a learned diagonal matrix.

    Second, it is unclear whether the advantage of relying on residual/subgradient oracles rather than constraint-specific projectors really materializes. To my best understanding, for SOC rows the subgradient and the cone projection are equally cheap. For the SDP, according to Appendix C each chordal PSD row still needs to compute a minimum-eigenpair, which for small chordal blocks has a cost about the same as the eigendecomposition behind LMI-Net’s projection. The QCQP is the only family where no simple projector exists, and there Π\Pi-net obtains the same gap by rewriting the constraints as SOCs (Table 5b). A constraint family where oracles are available but efficient projectors are not would make the generality claim by the authors more concrete.

    One more thing to point out here is that the method only provides ϵ\epsilon-feasibility since linearization halfspaces are outer approximations. In particular, for a curved convex constraint, the lineariztion halfspace contains the feasible set but is larger than it, so projecting onto the halfspace leaves the iterate slightly outside Ω\Omega. But then, this is the same type of guarantee as iterative repair methods such as DC3 and FSNet. The reported tolerances are also relatively loose (1e-3, and 1e-4 on QCQP), with the method sitting near the threshold in Figures 3(b) and 4(c) while Π\Pi-net reaches 1e-8 or better at comparable latency. Perhaps results at tighter tolerances would also be informative.

    Finally, the learned mobility is an input-dependent diagonal rescaling of the output coordinates, but given that all theory is for fixed PP, there is no analysis of how the learned geometry affects the convergence constants or the repair depth needed.

  2. The second main concern is that it is unclear at some parts whether the experiments really support the claimed contributions. First, the only ablation that evaluates the core inner solver (Table 6) shows dual PGD beats Violation-GD (essentially the DC3-style gradient correction), by just 1.06 to 1.17$\times$, while the large speed ups are over single-row repair and M-SSP, which are less competitive baselines. The closest competitors are also absent, e.g., T-SKM-Net on the affine task or learned active-set sampling (AT-SKM style) against greedy top-b selection. In addition, the evidence for learned mobility is also not strikingly convincing showing 0.245% to 0.190% five-seed mean on the affine task and 8 to 11% relative gains on SOCP300 and SDP300.

Questions for authors
  1. Can the authors clarify whether there is indeed a clear advantage of relying on residual/subgradient oracles over constraint-specific projectors? Could they bring up a constraint family where oracles are available but efficient projectors are not, and if so, could the method be demonstrated on one?
  2. How do latency and optimality gap change at tighter feasibility tolerances (e.g., 1e-5 or 1e-6)?
  3. How would the authors explain the small margin over Violation-GD in Table 6 combined with the poor performance of DC3 in Tables 2, 4 and 5?
2 comments
AuthorsOpenReview ↗

Responses to Reviewer DQL1

Thank you for the careful review and for suggesting an example where projection is costly.

What changes from T-SKM-Net? We build on T-SKM-Net's equality projection and null-space coordinates. Our update differs from a block Kaczmarz–Motzkin step in how it repairs several violations at once. We seek the smallest displacement satisfying the linearized inequalities in a metric learned with the predictor (Eq. 3). A selected inequality can remain slack. Both Equality-LS variants fit the selected boundaries. Neither reached tolerance on any of the five rotated-SOC instances. Median time to feasibility was 322.62 ms for Block-pinv and 2.36 ms for dual PGD (Table 24).

Under the assumptions of Cor. 4.2, the outer iterates converge even with one dual step per block, with a linear rate under the additional Slater condition.

In the full SOC-MPC pipeline, our layer took 0.94/1.06/3.13 ms at horizons 20/40/80, versus 76/153/304 ms for Π-net. At horizon 40 both met tolerance on every input (0.236% vs. 0.567% gap). On 300-bus SOCP, starting from the same predictor, our layer reached tolerance on 100% of inputs (41.7 ms on average) and FSNet on 8.5% (3,200 ms).

Oracle access (Q1). For log-sum-exp constraints, our layer uses the residual and its gradient, a softmax-weighted sum. Euclidean projection onto these sublevel sets generally requires a numerical solve. We computed it with Clarabel and CuClarabel. All three methods reached 10−610^{-6} on all nine instances (three per dimension).

DimensionOurs, RTX 4090 (ms)CuClarabel, RTX 4090 (ms)Clarabel, CPU (ms)
201.3654.983.81
502.4884.0020.27
1002.57160.90263.46

FP64 times include transfers and synchronization from CPU input to full CPU output. Our TorchAX/XLA repair and CuClarabel shared one RTX 4090. CuClarabel reused the fixed structure and reset its iterates. Mean squared displacement was 2.1–4.7% above CPU Clarabel's.

Tighter tolerances (Q2). As you point out, a halfspace step can leave a curved constraint violated. We stop when the maximum violation of the original constraints falls within tolerance. On QCQP (joint training, seed 456), both met 10−810^{-8} on all 833 inputs. Excluding transfers, dual PGD (power1) averaged 3.026 ms (0.01155% gap) and tolerance-stopped Π-net 7.671 ms (0.04013%).

All 5,120 affine seed–input pairs passed at each tolerance.

Tolerance10−310^{-3}10−510^{-5}10−610^{-6}10−810^{-8}
Mean latency (ms)0.6700.9381.0541.301
Mean gap (%)0.1590.1600.1600.160

In QCQP post-processing at 10−810^{-8} (inner and outer), ours took 0.82–0.93 ms over four seeds (833/833 each). Π-net took 23.1–23.6 ms at K=800K=800 with mean gaps within 0.0005 percentage points of ours. Stopping Π-net on tolerance took 0.68–0.70 ms with a 42.5–42.7% gap.

Mobility and convergence. Each input has its own P(x)P(x), fixed throughout repair. With a common Slater point and Euclidean subgradient bound, the upper bound on κ2\kappa_2 grows by at most a factor of cond(Pz)≤eℓ\sqrt{\mathrm{cond}(P_z)}\le e^\ell (Eq. 53). A step budget also requires the input's error-bound constant and initial distance to feasibility.

Table 6 and DC3 (Q3). From zero, Violation-GD (VGD) and dual PGD share the first step for matched rows and step size. Median paired time ratios (VGD/dual PGD) were 1.06–1.17 across six settings. We varied row correlation and half-width across nine affine settings. From identical starts, dual PGD reached 10−310^{-3} within 100 outer iterations on 461/576 cases and VGD on 384/576.

Tables 2, 4 and 5 compare full methods. We retrained the affine model with VGD using the same architecture, equality projection and row selection. Five-seed mean gaps were 0.172% for VGD (0.834 ms) and 0.159% for dual PGD (0.675 ms). VGD retained most of the advantage over DC3 (7.894% gap, Table 4).

VGD targets squared violation and dual PGD minimum displacement. On QCQP (seed 456), we trained one model per update. Both passed 833/833 at 10−810^{-8} using power1 at test time. Mean gaps were 0.01155% (dual PGD) and 0.01543% (VGD).

Affine baselines. T-SKM-Net passed all 5,120 pairs at a cap of 5,000 updates (0.528% gap, 2.40 ms). With a cap of 100 updates, 888 pairs missed tolerance. The b=64b=64 sweep selects all 50 affine rows. To test row selection, we used 36 rows (b=8b=8) and compared greedy top-8 with four greedy plus four learned rows, using the same predictor per seed. Both passed 320/320. Learning reduced selection rounds but raised mean time from 1.09 to 1.14 ms.

Mobility results. In Table 4 (b=16b=16), learned mobility reduced mean gap from 0.245% to 0.190% (four of five seeds improved). Separately trained b=64b=64 models gave 0.196% with identity and 0.159% with learned mobility at 10−310^{-3}. On SOCP300 and SDP300, learned mobility lowered gap versus head-only in both seeds. Mean latency fell on SOCP300 (35.7–36.4 to 24.4–29.9 ms) and rose on SDP300 (164–167 to 202–203 ms).

We hope these results address your concerns.

Thank you for the detailed responses and extra experiments.

My main concerns that remain are the following. First, dual PGD exceeds Violation-GD only marginally in the matched comparisons, which is something that indicates the large gains are coming from the surrounding pipeline rather than the from the block update. Second, the oracle advantage is demonstrated only on a small log-sum-exp test. Lastly, the provided evidence for learned mobility remains quite mixed.

Overall, the revision is worth acknowledging. However, I still see this work as an incremental extension of T-SKM-Net with the core component adding a limited advantage. Therefore I will maintain my weak reject score.

Comments on the paper

1 comment
Area_Chair_XQdDOpenReview ↗

Reminder: Follow-up Feedback Before Discussion Closes

Dear reviewers,

With the discussion period nearing its end, I would appreciate it if you could review the authors’ responses and revised manuscript and share their follow-up feedback.

Please indicate which concerns have been addressed and which remain, and update your assessment and score where appropriate.

Thank you for your time and contributions to the discussion.

Best, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.