acceptodds
Under review as a conference paper at ICLR 2027

Variational Surrogate Jacobians: A Framework for Constrained Gradient Propagation

Abstract

Many learning systems couple neural networks with discrete or constrained optimisation layers, but the resulting solution maps often have zero or undefined derivatives, preventing useful gradient propagation. We introduce the Variational Surrogate Jacobian (VSJ) framework, which selects the surrogate Jacobian closest to a prior backward rule under a positive-definite metric and has a unique closed form solution when selected active constraints are enforced exactly. To allow constraint enforcement to respond to the local optimisation state, we also introduce adaptive weighting of inequality constraints, which uses primal slacks and dual multipliers to set the enforcement strength. We show that VSJ recovers established backward rules as particular configurations under method-specific local assumptions, and also corrects those that depend directly on the loss. Empirically, VSJ closely reproduces six of them, while validation-selected novel VSJ configurations yield lower mean test regret for every tested prior family on at least one task. As an example, composing VSJ with PEAR, a state-of-the-art method, lowers mean Knapsack regret and increases mean Portfolio cumulative return. These results establish VSJ as a unified geometric framework for constructing, analysing, and designing gradient propagation rules through constrained optimisation layers.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
computer vision, deep learning, combinatorial optimization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_frUZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2(+1)
Presentation
2(+1)
Soundness
2
Read review
Paper summary

This paper introduces Variational Surrogate Jacobian (VSJ), a theoretically-motivated rule for how to backpropagate gradients through optimization-solution mappings, using a surrogate Jacobian, despite the actual mapping having zero or undefined gradients. The proposed class of surrogate Jacobians admits parameterizations which reproduce various other commonly used surrogate Jacobians, and acts as a generalized version. Sweeping over various hyperparameters (choice of M, Omega, and P) results in improved surrogate Jacobians, which when used for learning, increase benchmark scores on various optimization problems.

Critical strengths
  • Sweeping the possible VSJ configurations produces surrogate Jacobians with improved performance over baselines.
  • VSJ does not introduce a huge number of hyperparameters in order to be able to reproduce existing backward rules.
Critical weaknesses
  • Clarity: The paper is difficult to understand.
    • Proliferation of names and shorthands: By pages 8 and 9, there are too many terms lacking clear definitions which make the takeaways difficult or impossible to extract. Examples include "native" (line 361), "SPO+", "Blackbox", "cost reflection", "loss-gradient probing" (line 409), "loss-dependent updates" as a whole from lines 409 to 417, "starting recipe" (line 450), "matched strict-adaptive controls" and "native enforcement" (lines 450-451). The text explaining benchmark results is embedded with many numbers that are hard to contextualize.
    • Proliferation of symbols: New symbols are introduced at various points which are barely used elsewhere in the writing, rather than writing their definitions in-place directly, causing excess confusion, especially when the definitions are not present. Examples include QθQ_\theta, HentH_\text{ent}, and HlogH_\text{log} from Table 1, ee from line 303. It is unclear where τ\tau (line 296), ϵ\epsilon (line 300), and μLog\mu_\text{Log} (line 300) come from.
    • The problem setups of the benchmarks are underexplained and lack Figures. It is totally unclear from the main text what kind of optimization problem is set up from the benchmark description, how the learned predictor is used to pick the optimization parameters, and how the solution gets mapped to a loss, not to mention how a discrete solution is picked in the forward pass when the problem is a discrete optimization problem. Without these details, the many numbers embedded in the text in Section 5 lack context.
    • Motivations for various mathematical manipulations in Sections 1 and 3 are not clear before the manipulations are performed. E.g., Why do we choose to hard-enforce Fδs=0F \delta s = 0 for one part of the derivation (page 2, line 095) only to declare the hard-enforcement an issue (line 214) so to walk back on that decision in order to derive VSJ? It would be easier to follow if the assumptions required for deriving VSJ did not change as we derive it.
    • Lack of context on what RR function is used in various situations.
    • There is too much focus on the correctness of reproduction of other existing backward operators in the literature. It's not that hard to believe that the reproduction is correct, and the amount of tables, derivations, and numerical matching tolerances embedded in the text devoted to showing this correctness is distracting.
  • Alarming and unexplained mention of a "seed collapse" phenomenon on line 415, and also a mention that some ablations still remain to be tested (line 451).
  • There is no uniformly best selection of (P,M,Ω)(P, M, \Omega) for every problem, and the issue of picking the best algorithm to solve the benchmark is not reduced by the introduction of VSJ.
  • Line 419 instructs the practitioner that the practical method to retain settings from methods other than VSJ, and then tune hyperparameters.
  • Unlike OptNet (Amos & Kolter, 2017), VSJ does not provide gradients towards optimization problem components other than the objective, such as the constraint parameters.
Questions for authors
  • Why do different algorithms produce different results at all here? If I understand correctly, some of the benchmarks (Knapsack, Portfolio optimization problems) are under-parameterized, meaning with enough training, all converging algorithms using any Jacobians should all recover the same solution with the same loss, given that the feasible set and objectives are all convex. Are they fundamentally not guaranteed to all reach the same minimum in the end, or rather are the solutions under-trained, or is one of my assumptions wrong, such as the solution being discrete?
  • What is the "seed collapse" and why does it happen? How often does it happen?
7 comments

Response to Reviewer frUZ (1/4)

We sincerely thank Reviewer frUZ for the detailed, line-level feedback and for recognizing the improved backward configurations obtainable within VSJ’s compact parameterization. Your comments have guided revisions to terminology, notation, benchmark illustrations, and the motivation preceding the mathematical construction. We believe that the presentation of the paper improved quite notably. We also clarify few misunderstandings in our reply. We hope the revised manuscript, additional evidence and our further clarifications address your concerns. If so, we would be grateful for a reassessment. Please find our detailed responses below.

frUZ.1 Clarity concerns.

We thank the reviewer for the line-level detail, which made this weakness directly actionable. These concerns are not difficult to address, and doing so significantly improves the readability of the paper. Section 5 now defines every term at first use, and Section 4.1 now defines each symbol where it appears.

The revision addresses all points (clarity on names and shortands, the proliferation of symbols, the lack of benchmark figures, more motivations on equations). Importantly, none of results in the paper changes.

1. Proliferation of names and shorthands. Every term the reviewer lists is now defined at first use, and Section 5 now has a short terminology paragraph.

TermDefinition now in the paper
nativeA method's own backward rule as published, run from its official implementation where available, with no VSJ component. "Matched VSJ" is our single layer set to that method's row of Tab. 1.
SPO+The surrogate regret loss of Elmachtoub & Grigas (2022). Its gradient with respect to the predicted costs z is 2(s⋆(c) − s⋆(2z − c)), where c is the true cost vector.
BlackboxBlackbox differentiation (Pogančić et al., 2020). Its gradient is (s⋆(z + β ∂L/∂s) − s⋆(z))/β for a step size β > 0 (λ in the original paper).
cost reflectionOur shorthand for SPO+'s extra solver call at 2z − c, the true cost reflected through the prediction. Now defined in place by its formula.
loss-gradient probingOur shorthand for Blackbox's extra solver call at costs shifted along the loss gradient. Now defined in place likewise.
loss-dependent updateA backward rule that returns the cost gradient directly as a vector g, computed with one extra solver call that depends on the loss. A Jacobian-based rule instead returns Jᵀ ∂L/∂s for a matrix J fixed before the loss is seen.
starting recipeKeep the method's own P and M, and search only over Ω: none, strict, or adaptive.
matched strict-adaptive controlsThe "Strict VSJ" and "Adaptive VSJ" rows of Tab. 5. They share P, M, row selection and every hyperparameter, and differ only in Ω.
native enforcementThe enforcement the original method uses. For PEAR this is strict, Ω = I.

Also within this scope:

  • We added clarification sentences to the "Loss-dependent updates" paragraph of Section 5.3. Both SPO+ and Blackbox return a difference of two feasible decisions. Such a difference satisfies every equality the two decisions share, but the two decisions can lie on different faces of the feasible set.
  • A bound or capacity constraint that is tight at the current solution may therefore be slack at the other one. The update then has a component along that constraint's normal. Eq. (11) damps exactly that component, with strength set by Ω, and leaves an already tangent update unchanged.
  • The recipe now has one name, the "Ω-only recipe", and Sec. 5.4 points back to its definition. The sentence at lines 479-482 now reads: "The Strict VSJ and Adaptive VSJ rows of Tab. 5, which differ only in Ω, show why strict constraint enforcement should remain a candidate when configuring VSJ: adaptive enforcement lowers Knapsack mean regret from 1.9721% to 0.3834%, whereas strict enforcement gives lower Portfolio mean regret (81.25% against 81.88%)."
  • On the numbers embedded in the text, Section 5 now says how to read regret before reporting any. Zero means the decision is as good as the one made with the true costs, and a relative regret of 0.10 means an objective 10% worse. With that reading, each value in the prose maps to an entry of Tabs. 3–5, which also list its change against the native rule.

Response to Reviewer frUZ (2/4)

2. Proliferation of symbols

The reviewer points at three different problems, and we fixed each differently. We also added a new Notation section in appendix.

Symbols with no visible origin (τ at line 296; ε and µ_Log at line 300). None of these is a VSJ quantity. Each is the strength of the regulariser R in Eq. (1) that the corresponding existing method adds: τ for SparseMAP's quadratic term, ε for the entropy term, and µ_Log for the log barrier.

We had written those regularisers out only in Appendix F (Eqs. 105, 110, 113), so the main text showed their Hessians without their source. Tab. 1 now lists each regulariser beside its Hessian, so every symbol in the table is defined in the table:

MethodRegulariser R(s) in Eq. (1)Curvature H_R = ∇²R(s⋆)(P, M, Ω)
Identity Update00(−I, I, 0)
SparseMAP(τ/2)‖s‖²τI(−H_R⁻¹, H_R, I)
OptNet½ sᵀQ_θsQ_θ(−H_R⁻¹, H_R, I)
Entropy-KKTε Σᵢ sᵢ(log sᵢ − 1)H_Ent = ε diag(1/sᵢ⋆)(−H_R⁻¹, H_R, I)
Log-barrier KKT−µ_Log Σᵢ log sᵢH_Log = diag(µ_Log/(sᵢ⋆)²)(−H_R⁻¹, H_R, I)
PEAR (quadratic base)½ sᵀQ_θsQ_θ(−H_R⁻¹, H_R, I)

Here τ, ε and µ_Log are positive and are each method's own regularisation strength. Q_θ is the positive-definite quadratic-cost matrix of the QP, which OptNet may learn.

Symbols used once (e at line 303). The prediction error e appeared in a single sentence, together with g_PEAR and c. We removed all three from the main text and state the fact in words: PEAR's regret gradient applies the same Jacobian to the cost-prediction error. The derivation stays in Appendix F.

Symbols we kept (Q_θ, H_Ent, H_Log). We respectfully note that these are not single-use. They are the values of the prior and metric axes, and they label Fig. 1(c), Tab. 1, Tab. 3 and the appendix tables (Tabs. 20, 21 and 23–25).

3. Benchmarks are underexplained and lack Figures

We have now added Figure 3 to illustrate the five benchmark pipelines: input data → learnable predictor → predicted optimization inputs → evaluation solver → decision. Each panel specifies predictor dimensions, continuous or binary decision domains, and the objective and constraints. This complements:

  • Figure 1(a), which shows how decision outputs feed a downstream loss and how surrogate Jacobians enable backpropagation.
  • Appendix H.1 (Formulation of the Benchmarks), which provides the mathematical formulations of all five main benchmarks, including their optimization objectives, constraints, and continuous or binary decision domains.
  • Appendix H.2, which defines the evaluation metrics and explains how the resulting decisions are scored.

4. Why do we choose to hard-enforce F δs = 0 for one part of the derivation (page 2, line 095) only to declare the hard-enforcement an issue (line 214) so to walk back on that decision in order to derive VSJ? It would be easier to follow if the assumptions required for deriving VSJ did not change as we derive it.

We apologize for not making this progression clear. The assumptions in fact do not change during the derivation. To clarify, we first derive Strict VSJ, where FJ=0\mathbf F\mathbf J=\mathbf0 preserves the selected active manifold to first order. However, an inequality can remain satisfied without remaining tight, so this preservation can overly restrict updates. Adaptive VSJ therefore generalizes the construction by softening inequality enforcement in the backward rule, while keeping equalities exact. We now explain this motivation explicitly in the Introduction, before the detailed derivation in Section 3, to clarify that the Adaptive VSJ broadens the surrogate family rather than reverses an assumption.

We also included further clarification such as including Adaptive VSJ as a definition, and adding more intutions to equations.

5. Lack of context on what R function is used in various situations..

When we introduce R()R() function in Eq.(1) of Sec.1, we include more intuition and examples. Specifically, we mention that "Setting R=0R=0 gives a linear objective, while quadratic (Amos & Kolter, 2017) or entropic (Agrawal et al., 2019) choices give the regularized problems considered later."

Response to Reviewer frUZ (3/4)

6. Too much focus on the correctness of reproduction of other existing backward operators

We would like to kindly clarify that these experiments (in Section 5.2) are not merely reproductions of baseline performance. VSJ provides a single variational construction in which existing backward operators arise as particular choices of the prior, metric, and constraint weights. Section 5.2 therefore checks whether selecting those configurations in the unified VSJ implementation recovers the corresponding native operators, examining both operator agreement at matched problem states and end-to-end consistency. Therefore, this directly supports the framework’s unification claim and provides a controlled reference for evaluating new configurations.

frUZ.2 What is the "seed collapse" and why does it happen? How often does it happen?

What is collapse? Intuitively, the convolutional network (that is supposed to learn the mapping from the country flags to their locations on the globe), maps different countries into an almost point-like configuration, instead of learning their relative locations. This is a previously documented difficulty, not a failure introduced by VSJ. For example,

How often does it occur? To quantify this behaviour and investigate its mechanism, we extended the TSP comparison from 5 to 20 paired seeds, evaluating SPO+ and Blackbox with and without their selected VSJ configurations. We retain every seed and distinguish three observed outcomes by the fraction of test instances decoded to an optimal tour:

  • Collapsed: 0% optimal tour predictions from the model.
  • Stalled: >0% but less than 50% optimal tour predictions.
  • Regular: at least 95% optimal tour predictions.

Collapse is therefore the more severe failure: no test instance receives an optimal tour, rather than training attaining only partial success.

VSJ substantially reduces the observed failure frequency: We have noted that no collapse occurs with either selected VSJ configuration, and SPO+ stalling decreases from 11 to 5 runs.

MethodCollapsedStalledTotal unstable runs
Native Blackbox4/20 (20%)0/204/20 (20%)
Blackbox + VSJ0/201/20 (5%)1/20 (5%)
Native SPO+0/2011/20 (55%)11/20 (55%)
SPO+ + VSJ0/205/20 (25%)5/20 (25%)

Does this improve performance? Yes. Across all 20 seeds, mean relative test regret decreases (lower-better):

  • Blackbox: from 0.4549±0.93340.4549\pm0.9334 to 0.0030±0.01350.0030\pm0.0135.
  • SPO+: from 0.0133±0.01480.0133\pm0.0148 to 0.0065±0.01290.0065\pm0.0129.

Thus, improved training reliability is reflected in better performance.

How does this relate to the theory? As discussed in Section 4.2, differences of feasible decisions preserve shared affine equalities but need not be tangent to selected active inequality or bound constraints. VSJ controls their enforcement through Ω\mathbf\Omega and the correction geometry through M\mathbf M. The observed reductions in collapse and stalling, together with lower all-seed regret, provide empirical evidence of improved training reliability with the selected VSJ configurations on this benchmark, consistent with this motivation.

Manuscript change: Having provided an in-depth analysis, we now report the main-paper TSP comparison over 20 seeds without excluding unsuccessful runs for loss-dependent updates. We have also added a detailed discussion in Appendix H.9 for the collapsed and stalled cases.

frUZ.3 — Apparent unfinished ablations

We apologize for the ambiguous wording. “Native enforcement remains worth testing” refers to the experiment we had already completed and included in the initial submission to highlight Strict VSJ's performance in Table 5. With this sentence, we wanted to emphasize that Strict VSJ, in fact, performs better than Adaptive VSJ for time-series Portfolio in Table 5. To clarify, we have now replaced that sentence with: “The Strict VSJ and Adaptive VSJ rows of Tab. 5, which differ only in Ω\Omega, show why strict constraint enforcement should remain a candidate when configuring VSJ.”

Reviewer_vqMnOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies backpropagation through discrete or constrained optimization layers, where the solution map has zero or undefined derivatives. Traditional approaches utilize a surrogate Jacobian for training. The authors propose the Variational Surrogate Jacobian (VSJ), which selects the surrogate Jacobian closest to a prior backward rule under a positive-definite metric, subject to preserving the active constraints to first order. This yields a closed-form solution that projects the prior onto the tangent space of the active constraints, and separates backward rule design into three choices: the prior PP, the metric MM, and the constraint weighting Ω\Omega. They further introduce adaptive weighting, which sets the enforcement strength of each inequality from its primal slack and dual multiplier. They show that several existing backward rules are recovered as particular VSJ configurations and that VSJ can also correct loss-dependent updates such as SPO+. Experiments on five benchmarks, including TSP, Knapsack, and Portfolio optimization, show that validation-selected VSJ configurations reduce test regret compared with existing rules, and that combining VSJ with PEAR further improves performance.

Critical strengths

1. Unifying view of backward rules

By placing existing methods in a common (P,M,Ω)(P, M, \Omega) space, the paper makes explicit which design choices each method couples together. Decoupling these choices opens a design space of new backward rules that previous methods could not express.

2. Improvement over a strong baseline

Combining VSJ with PEAR reduces regret on binary Knapsack, and adaptive VSJ appears to avoid the training collapse observed for some native methods on TSP.

Critical weaknesses

1. Experimental gains

It is unclear whether the improvement comes from the proposed constraint correction or from other factors. In Table 4, several selected configurations use Ω=0\Omega = 0, which means no correction at all. In these cases, performance improves by removing the native correction rather than by adding VSJ’s correction.

2. Applicability to discrete problems

The geometric motivation of VSJ does not directly hold for discrete problems. For TSP, the active constraints and dual multipliers come from an LP relaxation, not from the actual discrete solution.

Questions for authors
  1. For configurations selected with Ω=0\Omega = 0, how should the gains be attributed to VSJ? How do native methods perform when given the same tuning budget as VSJ?
  2. For discrete problems, how much does the performance depend on using the LP relaxation to select constraints and compute dual multipliers? How does VSJ behave when the relaxation is far from the discrete solution?
  3. How sensitive is adaptive VSJ to μ\mu?
  4. What is the computational overhead of VSJ?
6 comments

Response to Reviewer vqMn (1/4)

We sincerely thank Reviewer vqMn for the constructive review and for recognizing VSJ’s unifying view of backward rules and its improvements over the strong PEAR baseline. Your questions motivated substantial new evidence: controlled constraint-correction ablations, a matched-budget comparison with dXPP, relaxed–integer constraint-agreement diagnostics, temperature-sensitivity experiments, and expanded computational profiling. We hope these additions address your concerns and support a more favorable assessment. Please find our detailed responses below.

vqMn.1 Attribution of improvements to constraint correction, including the cases with Ω=0\Omega=0. How do native methods perform when given the same tuning budget as VSJ?

VSJ separates the prior response PP, correction geometry MM, and constraint enforcement Ω\Omega, allowing these choices to be examined independently. This includes disabling correction, giving J⋆=PJ^\star=P, rather than assuming that constraint enforcement always improves learning. We distinguish evidence for adaptive correction from gains obtained by removing correction below.

  • The benefit of constraint correction. To isolate the benefit of adaptive constraint correction, we added a binary-Knapsack ablation comparing adaptive enforcement with (i) strict enforcement, (ii) shuffled adaptive weights, and (iii) no correction. Across these VSJ variants, we hold the prior, metric, forward solver, per-seed data and initialization, and training protocol fixed. We additionally compare with dXPP (Linghu et al., ICML 2026), a very recent work which also uses primal–dual information for soft constraint enforcement in the backward pass. It combines slack-dependent penalty curvature with a shared multiplier-derived scale, whereas adaptive VSJ uses each inequality’s individual multiplier and slack. This makes dXPP a relevant comparison against an existing state-dependent soft-enforcement method, beyond the strict-enforcement and no-correction controls. The following table shows that Adaptive VSJ achieves lower regret than each comparator in all 10 paired seeds. Its improvement over Ω=0\Omega=0 demonstrates that the gains in this setting do not arise from removing correction. The shuffled-weight control preserves the weight values within each constraint type at a given state but changes their assignment to constraints. Its deterioration supports the importance of constraint-specific weight assignment, rather than simply applying nonzero correction.
MethodTest regret (%) ↓
Adaptive VSJ0.3834±0.02840.3834\pm0.0284
dXPP (Linghu et al., 2026)0.4587±0.03460.4587\pm0.0346
Strict VSJ1.9721±1.25251.9721\pm1.2525
Shuffled adaptive weights2.4204±0.29472.4204\pm0.2947
No correction (Ω=0\Omega=0)3.1431±0.89133.1431\pm0.8913
  • Interpreting the Ω=0\Omega=0 cases. VSJ separates the prior response PP, correction geometry MM, and constraint enforcement Ω\Omega, allowing these choices to be examined independently. Accordingly, setting Ω=0\Omega=0 gives J⋆=PJ^\star=P, providing a prior-only control for each prior, a configuration not used by the considered approaches other than Identity Update. Three validation-selected configurations in Table 3, using Entropy-KKT or Log-barrier-KKT priors, take this form. These cases illustrate the value of VSJ’s modular design: it allows validation to identify when retaining the prior response without correction is preferable. The gains in this section arise from removing correction, rather than from adaptive constraint enforcement.

Response to Reviewer vqMn (2/4)

  • How do native methods perform when given the same tuning budget as VSJ? Thanks for raising this. We used established benchmark implementations and published training protocols, where possible. And otherwise (e.g., applying a method to a new benchmark, hence no prior reference), we generally provided similar budgets to the native approach with VSJ. More specifically,
    • PEAR (Lee et al., 2026), Table 5. We use the official PEAR implementation and its recommended settings. Our reproduction closely matches its published binary-Knapsack performance and produces slightly better results for time-series Portfolio.
    • dXPP (Linghu et. al, 2026) in Figure 4(a,b) As dXPP was not applied to binary knapsack task, we conduct a matched 12-configuration search per method, using learning rates (0.003,0.01,0.03\\{0.003,0.01,0.03\\}) for both and μ∈0.001,0.01,0.1,1\mu\in\\{0.001,0.01,0.1,1\\} for Adaptive VSJ and (δ∈10−6,0.01\delta\in\\{10^{-6},0.01\\}) and (ζ∈1,10\zeta\in\\{1,10\\}) for dXPP. Both select their configuration by mean validation regret over the same five seeds. Thus, this restricted comparison matches both the number of candidate configurations and the number of validation seeds.
    • PredOpt-derived benchmarks (Mandi et al., 2024), Tables 3–4. We inherit benchmark implementations and training settings where applicable. Mandi et al. select method-specific settings through validation grid search over eight learning rates and the relevant method hyperparameters. Hence, the native approaches are already tuned extensively by Mandi et al. While we include VSJ, we simply borrow the settings of the native approaches (including learning rate), hence do not further tune each VSJ configuration. Within this scope, we have carefully checked our experiments again, and noticed 2 mismatches in our settings compared to PredOpt (SparseMAP and Log-Barrier for Fractional Knapsack in Table 3). Aligning them has not affected our major conclusions.
    • TSP (Pogančić et al., 2020), Tables 3–4. Each VSJ configuration uses the inherited benchmark training hyperparameters, including the learning rate, without separately retuning them. We select the VSJ backward configuration on validation.

Manuscript change: Section 5.5 now reports the constraint-weight ablation and dXPP comparison, and distinguishes gains from adaptive correction from those obtained by removing correction.

Response to Reviewer vqMn (3/4)

vqMn.2 Applicability to discrete problems and dependence on LP-relaxation quality.

Thank you for raising this point. First, we would like to mention that using continuous relaxations for backward surrogates is shared with all existing relaxation-based native approaches (such as PEAR (Lee et.al, 2026), Log-barrier-KKT (Mandi & Guns, 2020)) unified by VSJ. So, it is not unique to our method.

We agree that relaxation-derived constraints describe a surrogate geometry, not the active manifold of an integer solution. To assess whether this geometry retains relevant information, we have now examined agreement of active (tight) constraints: whether the relaxed and integer solutions reach the same constraint boundaries, such as an item's lower or upper bound. This is directly relevant because these constraint normals enter VSJ's backward correction; similar objective values alone would not establish such agreement.

Specifically, on binary Knapsack, we compare the smoothed QP state actually used by Adaptive VSJ-PEAR with exact binary solutions at the same predicted costs. We evaluate 100 training instances per seed over ten seeds at three training stages. Across these stages, on average across seeds, 97.2-98.5% of the constraints active in the smoothed state are also active at the integer solution, and the smoothed state recovers 90.2-91.5% of the integer solution's active constraints.

For TSP, separate diagnostics at Blackbox+VSJ checkpoints show that, on average across seeds, 95.5-96.1% of edges assigned value one by the subtour-strengthened LP belong to the exact tour, recovering 85.8-87.9% of the tour's edges. An edge at value one is fully selected and lies on its upper bound, so these statistics measure agreement of active upper-bound constraints, similar to what we did for Binary Knapsack. The first percentage says that almost all fully selected LP edges are also in the exact tour, and the second says that most exact-tour edges are already fully selected by the LP, rather than fractional or unselected.

These measurements show substantial active-constraint agreement for the actual Knapsack backward state and upper-bound agreement for the diagnostic TSP relaxation, even when relaxed and integer solutions differ. This supports the relevance of relaxation-derived constraint information in the examined discrete settings, providing motivation for its use in VSJ and related relaxation-based approaches.

Manuscript change: We have now added Active-Constraint Agreement Between Relaxed and Integer Solutions as Appendix H.12 to the paper.

vqMn.3 How sensitive is adaptive VSJ to μ\mu?

For fixed primal–dual information, a smaller temperature μ\mu strengthens inequality enforcement through γ=λ/(r+μ)\gamma=\lambda/(r+\mu), while a larger μ\mu weakens it. To address the reviewer's concern, we have examined the sensitivity of Adaptive VSJ to μ\mu across a broad temperature range on the binary Knapsack task using two reference methods: (i) Strict VSJ, a hard-enforcement reference; and (ii) dXPP, a recent primal–dual-based soft-enforcement baseline.

Specifically, we vary the base temperature μ∈0.001,0.01,0.1,1,10\mu\in\\{0.001,0.01,0.1,1,10\\}, with μbound=μ\mu_{\mathrm{bound}}=\mu and μcapacity=10μ\mu_{\mathrm{capacity}}=10\mu, keeping the remaining settings fixed across the VSJ variants. dXPP is separately validation-tuned. Results are mean test regret ±\pm sample standard deviation over 10 paired seeds.

Methodμ\muTest regret (%) ↓
dXPP (Linghu et al., 2026)N/A0.4587±0.03460.4587\pm0.0346
Strict VSJN/A1.9721±1.25251.9721\pm1.2525
Adaptive VSJ0.0010.0010.4558±0.24590.4558\pm0.2459
Adaptive VSJ0.010.010.3834±0.02840.3834\pm0.0284
Adaptive VSJ0.10.10.4502±0.04060.4502\pm0.0406
Adaptive VSJ110.5521±0.05610.5521\pm0.0561
Adaptive VSJ10100.6746±0.07070.6746\pm0.0707

Adaptive VSJ achieves lower mean test regret than Strict VSJ at every tested temperature, spanning four orders of magnitude. At the three tested values μ∈0.001,0.01,0.1\mu\in\\{0.001,0.01,0.1\\}, its mean regret is also lower than dXPP's. The validation-selected μ=0.01\mu=0.01 achieves the lowest mean regret, while larger tested temperatures progressively increase mean regret. Thus, μ\mu matters for achieving the best performance, but the mean-regret advantage over strict enforcement persists throughout the tested grid in this setting.

Manuscript change: We have added this temperature-sensitivity analysis and its accompanying figure (the new Fig.4b) to the new ablation study section in Sec. 5.5 of the revised manuscript.

Reviewer_Fm8xOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The authors organize several established constrained gradient propagation methods into a unified framework named Variational Surrogate Jacobian. They use a surrogate Jacobian for propagating the gradients through an optimization layer. This surrogate is defined as the closed-form solution to a quadratic form, with 3 hyper-parameters: the prior PP, the metric MM and the weighting Ω\Omega. They show that several established constrained gradient propagation methods correspond to particular choices of (P,M,Ω)(P, M, \Omega), and find novel choices that outperforms these baselines.

Critical strengths

The authors provide rigorous geometrical interpretation for the configuration space of (P,M,Ω)(P, M, \Omega) and show the explicit configurations required to replicate each of the established baselines. They provide empirical evidence that these configurations replicate the baselines numerically. The proposed framework allows the search for new configurations that has shown improvements over the state-of-the-art on established benchmarks, namely:

  • The Traveling Salesman Problem (TSP)
  • Fractional Knapsack
  • Binary Knapsack
  • Time-series Portfolio
Critical weaknesses

The results shown on the tables would benefit from a larger pool of seeds. The confidence intervals represent only 1 standard deviation, and some of the most important results are intersecting. For example, the reduction in mean binary knapsack regret by 13.8%:

  • PEAR: 0.4358 ± 0.0439
  • Adaptive VSJ-PEAR: 0.3756 ± 0.0312

This is an improvement suggestions and does not invalidate the results, since the VSJ method wins on several benchmarks.

Questions for authors

What would be the impact of using non-linear objective functions?

1 comment

Response to Reviewer Fm8x

We sincerely thank Reviewer Fm8x for the positive assessment and for recognizing the rigorous geometric interpretation, numerical recovery of established operators, and improvements over strong baselines. Your suggestion to increase the seed count prompted us to extend both PEAR-comparison tasks to 10 seeds, preserving mean-regret improvements of 15.1% on binary Knapsack and 2.6% on time-series Portfolio. We also clarify how nonlinear objectives are already represented in our formulation and experiments. We hope this expanded evidence and clarification address your questions. Please find our detailed responses below.

Fm8x.1 The tables would benefit from a larger pool of seeds.

We have now extended all methods in the PEAR-comparison table to 10 seeds for both binary Knapsack and time-series Portfolio. Adaptive VSJ–PEAR’s mean-regret improvement over PEAR is preserved on both tasks:

  • Binary Knapsack: mean test regret is reduced by 15.1% over 10 seeds, compared with 13.8% over the original five seeds. Adaptive VSJ–PEAR outperforms PEAR in all 10 paired seeds.
  • Time-series Portfolio: mean test regret is reduced by 2.6% over 10 seeds, compared with 2.3% over the original five seeds.

Results below are test regret (%), reported as mean ±\pm sample standard deviation.

BenchmarkMethod5 seeds10 seeds
Binary KnapsackPEAR0.4358±0.04390.4358\pm0.04390.4436±0.04660.4436\pm0.0466
Binary KnapsackAdaptive VSJ–PEAR0.3756±0.03120.3756\pm0.03120.3766±0.02930.3766\pm0.0293
Time-series PortfolioPEAR82.41±4.6882.41\pm4.6884.86±5.8984.86\pm5.89
Time-series PortfolioAdaptive VSJ–PEAR80.48±3.7780.48\pm3.7782.63±5.1482.63\pm5.14

Manuscript change: We have updated all binary-Knapsack experiments and the time-series Portfolio comparison to include 10 seeds. We will also aim to extend the remaining experimental results to more seeds for the camera-ready version.

Fm8x.2 The impact of using non-linear objective functions:

To provide a complete answer on the effect of non-linearity, we distinguish nonlinearity in the objective from nonlinearity in the constraints below:

  • 1. Nonlinearity in the objective. VSJ is not restricted to objectives linear in the decision variables: Eq. (1) already allows z⊤s+R(s)z^\top s+R(s), with smooth nonlinear RR, including quadratic, entropy, and barrier terms. Our time-series portfolio experiment (Fig.3(e) - real world data) uses the quadratic mean–variance objective. max⁡s;z⊤s−λ2s⊤Σ^s.\max_s\\;z^\top s-\frac{\lambda}{2}s^\top\widehat{\Sigma}s.
    The backward rule differentiates with respect to predicted returns while holding Σ^\widehat{\Sigma} fixed. Table 5 reports lower mean test regret and higher mean cumulative return for Adaptive VSJ–PEAR than for PEAR, providing empirical evidence with a nonlinear objective.

  • 2. Nonlinearity in the constraints. VSJ also accommodates smooth nonlinear constraints through their local Jacobians. Theorem 1 characterizes first-order preservation of the selected constraints. Intuitively, a smooth curved constraint boundary is approximately flat locally, so its tangent geometry can guide small, boundary-preserving decision perturbations. However, curvature limits this approximation for larger movements, and strict tangency can suppress useful changes in which constraints are active. To assess practical performance, we also include the synthetic portfolio experiment with the quadratic risk cap (s⊤Σs≤γs^\top\Sigma s\le\gamma). In Table 3, validation-selected VSJs lower mean regret for the Entropy-KKT and Log-barrier-KKT priors. Similarly, Table 4 shows that nonzero VSJ corrections improve both SPO+ and Blackbox on this nonlinear-constraint task.

Thus, nonlinear objectives are already within our formulation and experiments. For nonlinear constraints, we distinguish local surrogate guarantees from exact solution sensitivities and global guarantees.

Comments on the paper

1 comment

General Comment

We sincerely thank all reviewers for their constructive feedback and time. We appreciate their recognition of VSJ’s rigorous geometric interpretation and numerical recovery of established operators (Fm8x), its unifying perspective and independently configurable design choices (vqMn), and its empirical improvements over existing methods (Fm8x, vqMn, frUZ).

We would like to highlight that VSJ’s central contribution is a configuration space for backward propagation: separating the prior response, correction geometry, and constraint enforcement allows established operators to be recovered under their local assumptions and their components to be varied independently, enabling new configurations with improved empirical performance. In response to the reviews, we have strengthened the evidence for this contribution and substantially clarified its presentation. Below is a summary of the main additions and clarifications:

  • Stronger multi-seed evidence. We extended the PEAR comparisons to 10 seeds on both binary Knapsack and time-series Portfolio. Adaptive VSJ-PEAR preserves its mean-regret improvements: 15.1% on binary Knapsack, outperforming PEAR in all 10 paired seeds, and 2.6% on time-series Portfolio. These results strengthen the evidence beyond the original five-seed comparisons. (Fm8x.1)

  • Further ablations of Adaptive VSJ and comparison with dXPP (Linghu et al., ICML 2026), a recent baseline for soft constraint enforcement. We strengthened the evaluation in three complementary directions:

    • Controlled constraint-correction ablations. A new ten-seed binary-Knapsack ablation holds the prior, metric, forward solver, and training protocol fixed while comparing adaptive, strict, shuffled-weight, and no-correction variants. Adaptive VSJ outperforms each control in all 10 paired seeds, supporting the value of constraint-specific enforcement. (vqMn.1)
    • Comparison with dXPP. Using matched 12-configuration validation searches, Adaptive VSJ achieves 0.3834% versus 0.4587% mean test regret, extending the comparison beyond strict-enforcement and no-correction controls to a recent state-dependent soft-enforcement method. (vqMn.1)
    • Temperature sensitivity. We evaluate Adaptive VSJ across five temperatures spanning four orders of magnitude. Although temperature selection affects performance, Adaptive VSJ retains lower mean regret than Strict VSJ at every tested value on binary Knapsack. (vqMn.3)
  • Clarifications of ambiguities, motivation, and scope.

    • Clarifications of ambiguities and motivation. We explain how adaptive enforcement generalizes the strict construction and why recovering established operators is central to VSJ’s unification claim. We also clarify that the ablations described as “worth testing” were already completed. For TSP, we contextualize collapse as a previously documented difficulty and distinguish it from stalling. Across 20 paired seeds, retaining every run, the selected VSJ configurations reduce unstable runs from 4/20 to 1/20 for Blackbox and from 11/20 to 5/20 for SPO+, alongside lower all-seed mean regret. (frUZ.1–3)
    • Clarifications of scope. New relaxed–integer diagnostics support the relevance of relaxation-derived constraint information in the examined discrete settings, without equating it with integer-solution geometry. We clarify how nonlinear objectives are already covered, why convex inner problems need not yield convex end-to-end learning, and that our operator recoveries concern objective-input sensitivities; differentiation with respect to constraint parameters remains future work. (Fm8x.2; vqMn.2; frUZ.4–6)
  • Improved presentation. We further clarified abbreviations, terminology, symbols, and regularizer choices, and added Figure 3 to illustrate the five main benchmark pipelines, making the connections between inputs, predictors, optimization problems, and decisions explicit. We also strengthened the motivation preceding the mathematical construction. (frUZ.1)

Detailed point-by-point responses are provided below.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.