acceptodds
Under review as a conference paper at ICLR 2027

Controlling Scale Mismatch in Decision-Focused Learning with Cost Regularization

Abstract

Decision-focused learning (DFL) trains predictive models to minimize the cost of the decisions they induce rather than prediction error. When the model predicts the objective coefficients of a combinatorial optimization problem, the map from costs to decisions is piecewise constant, and many state-of-the-art methods obtain informative gradients by perturbing the predicted costs. We show that these methods share a common weakness: the perturbation scale is fixed, but the scale of the predicted costs is not. Because decisions are unchanged when costs are rescaled, the training loss does not constrain that scale, and it can drift freely during training. When predicted costs grow too large relative to the perturbations, the perturbed decisions stop differing from the unperturbed ones, and learning stalls or reduces to imitation. When they become too small, the perturbations dominate and the gradients no longer reflect the prediction. Since this drift varies across samples and over training, tuning the perturbation scale alone cannot prevent it. We propose a simple, method-agnostic remedy: a smooth shrinkage of the predicted costs, which leaves decisions unchanged and adds negligible cost. Experiments with three DFL methods show that it markedly reduces regret on a problem designed to induce this instability, and on standard benchmarks it corrects severe training failures at little cost when none occur.

Primary area:
optimization
Keywords:
decision-focused learning, empirical cost minimization, optimization, machine learning

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_xxe5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

To address an important issue in perturbation-based decision-focused learning (DFL), this paper proposes a shrinkage approach that rescales the predicted coefficients. Experiments demonstrate that the proposed method improves regret across existing perturbation-based DFL approaches.

Critical strengths
  1. The paper addresses a key challenge in perturbation-based decision-focused learning: the perturbation scale is fixed, whereas the scale of the predicted costs is not, which harms the performance of erturbation-based DFL approaches. The paper develops a natural approach to this issue by shrinking the predicted coefficients.
  2. The motivation is well-founded, and the theoretical analysis is sufficient to support the proposed approach.
Critical weaknesses
  1. The paper does not clearly state the research scope in the begining part of the paper. In the introduction, the authors state that "the map from costs to decisions is piecewise constant." However, this property holds only for combinatorial optimization problems rather than the general optimization problems. The paper's title and abstract also do not explicitly specify this setting, although it is clarified in the later text.
  2. The introduction states that the primary focus is "the problem of estimating objective function costs in combinatorial optimization problems." However, I do not think this accurately captures the broader scope of decision-focused learning. First, DFL is not restricted to combinatorial optimization problems. Second, estimating objective function costs is one approach to handling uncertainty in optimization problems (e.g., uncertain coefficients) and achive better decision quality. However, estimating objective function costs is not the only way for DFL. The paper should better clarify how its problem setting relates to the broader DFL literature.
  3. The introduction lacks clarity, especially for readers who are not already familiar with this research area, they may struggle to understand the underlying challenge and the motivation for the method.
  4. Some notation appears to be missing or insufficiently defined. For example, the notation in Equation (6) needs clarification.
Questions for authors
  1. How is κ\kappa selected? In practice, does achieving good performance require an exhaustive hyperparameter search over κ\kappa? If so, how sensitive is the method to this choice?
  2. The evaluation appears to be limited to a three-vertex toy problem. How does the proposed regularization perform on larger-scale mixed-integer linear programming (MILP) problems or graph optimization tasks? Furthermore, does the optimal value of κ\kappa change systematically as the number of vertices increases?
  3. The proposed regularization applies a forward mapping that directly transforms θ\boldsymbol{\theta}, rather than introducing a penalty term into the objective function (e.g., an L2L_2 regularization term). What is the motivation for choosing this transformation over a conventional penalty such as λ∥θ∥22\lambda\|\boldsymbol{\theta}\|_2^2? Could the authors clarify the theoretical or empirical advantages of the proposed approach over standard regularization?
2 comments
AuthorsOpenReview ↗

Answers to Reviewer 3.

We thank the reviewer for their comments and questions.

W1. We broadened the definition of decision-focused learning in the abstract, and we specified that our study focuses on cost estimation for combinatorial optimization.

W2. We broadened the definition of decision-focused learning in the introduction, and we specify that our study is restricted to cost prediction for Mixed-Integer Linear Programs. Additionally, a paragraph was added at the beginning of the literature review to present the broader scope of DFL applications.

W3. Beyond the specific changes mentioned above, the introductory sections have been revised for clarity.

W4. We clarified the notation. In particular, there was a typo in Equation 6: f~\tilde f was replaced by ff (f~\tilde f denotes a perturbed decision mapping, required by another version of the Fenchel-Young loss). Moreover, we now introduce the loss LFYL^{FY} before it appears in Equation 6.

Q1. κ\kappa is selected on validation during hyperparameter optimization, so no exhaustive search is required. As mentioned in Section 5, κ\kappa balances flexibility and regularization: larger values yield looser bounds on the stability radius. Its best value is therefore problem-specific, a compromise between flexibility during training and ensuring that perturbations and cost estimates remain on the same scale. Regarding sensitivity, the extremes of κ\kappa are known methods: as κ→∞\kappa \to \infty the shrinkage falls back to the standard method, and as κ→0\kappa \to 0 it approaches normalization, so a poor choice drifts towards one of these baselines rather than causing a collapse (see also our answer to Reviewer 2).

Q2. Beyond the three-vertex problem, we now evaluate our approach on two standard benchmarks of the PyEPO library: shortest path on a 5×55\times5 grid (40 edges) and a 32-item knapsack, with five degrees of nonlinearity of the feature-to-cost mapping (Section 6.2, Appendix E). We sshow that the cost regularizations outperform standard learning in the cases expected to exhibit instability.

Q3. Our new experiments now compare the L2L_2 penalty to the standard and regularized methods. Although the L2L_2 penalty outperforms other approaches in a few cases (DPO in c1 and c2 of the three-vertex problem), it is inconsistent: it is worse than the standard method for DBB on every benchmark and for SPO in c3, and it falls far behind clipping for DPO in c3, where instability is strongest. Overall, cost regularization is the more reliable alternative: it guarantees a bound on the norm of the cost fed to the perturbation, whereas the L2L_2 penalty may require much more use-case-specific tuning to reach comparable results.

Dear Reviewer, as the discussion period ends on November 18, we would be grateful if you could let us know whether our response addresses your concerns. In particular, we clarified the scope of the paper and mpared our approach with an explicit L2 penalty as you suggested. We also note that, beyond the three-vertex problem, the paper now evaluates the approach on new realistic experiments to demonstrate its relevance, namely the standard shortest-path and knapsack benchmarks (Section 6.2). We are happy to clarify any remaining point.

Reviewer_Psh1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper is written in the context of perturbation-based Decision-Focused Learning (DFL) methods. Existing methods can result in unstable or stalled training because of the drifting scale of cost estimates relative to perturbations. The paper provides a result showing that, for a fixed direction, the stability radius of a cost vector is proportional to its norm, thus requiring regularization of cost vector estimates. The paper then proposes a smooth shrinkage regularization that bounds the norm of predicted cost vectors while preserving their optimal decisions. On three combinatorial optimization tasks, the proposed method reduces regret and mitigates severe training failures compared with unregularized SPO, DPO, and DBB methods.

Critical strengths

The paper studies an important shortcoming in existing works in DFL : training collapse in perturbation-based DFL due to solution stability and solution invariance.

The insight that without controlling solution stability, DFL techniques designed for learning through experience become either inefficient or degenerate into imitation-based learning is especially useful for understanding when these methods lose their ability to learn from the costs of downstream decisions.

Critical weaknesses

While the paper compares the proposed regularization against unregularized methods and L2 normalization, it does not consider other simple baselines to fix the issue, such as clipping predicted cost vectors to a maximum norm or adding an explicit penalty on their norms. These comparisons would help establish the benefits of the proposed smooth shrinkage.

The numerical experiments section is hard to understand and does not clearly describe the benchmark tasks and instance categories on which the methods are compared.This hinders interpretation of the results.

Questions for authors

The paper mentions that “It is commonly assumed that the optimization mapping f is piecewise constant and thus possesses a gradient that is of little relevance in an optimization process”. Can you elaborate more on why this mapping is piecewise constant?

How sensitive is the proposed method to the regularization parameter k?

The paper mentions that degeneration can lead to imitation based learning. Can the authors clarify when degeneration into imitation-based learning leads to worse regret?

3 comments
AuthorsOpenReview ↗

Answers to Reviewer 2.

We thank the reviewer for their positive assessment and helpful suggestions.

W1 (baselines). We now compare our approach to two additional baselines, norm clipping and an explicit L2L_2 penalty on the predicted costs, in new experiments on standard benchmarks. Clipping performs close to the shrinkage, which confirms that bounding the scale of the cost fed to the perturbation is what matters; the shrinkage is the smooth member of this family. The L2L_2 penalty generally performs worse than the cost regularizations.

Overall, scale mismatch arises on standard benchmarks, even with normalized targets, and grows with the heterogeneity of the cost scale. Bounding the norm of the cost fed to the perturbation works best: the shrinkage rsr^s, the safer default, or clipping rcr^c, stronger under extreme instability, improves DPO everywhere, SPO on the three-vertex problem and DBB on most benchmark settings. The normalization rnr^n fails on small cost vectors, the L2L_2 penalty lags where instability is strongest, and target normalization does not prevent the collapse.

W2 (clarity of experiments). The benchmark tasks, data generation, instance sizes and models are now fully described in Appendix E, and the three-vertex problem in Appendix D.

Q1 (piecewise constant mapping). We improved the scope definition in the early sections regarding the focus on combinatorial optimization problems. For a linear objective over a finite set YY, the cost space is partitioned into the normal cones N(y)={θ:θ⊤y≤θ⊤y′ ∀y′∈Y}N(y) = \{\theta : \theta^\top y \le \theta^\top y' \ \forall y' \in Y\}. The mapping ff is constant on the interior of each cone and can only change on their boundaries, which have measure zero. Hence ∇f=0\nabla f = 0 almost everywhere and is undefined on the boundaries. This is a property of the problem, not an assumption, and it is exactly what the stability radius quantifies: η(θ)\eta(\theta) is the distance from θ\theta to the boundary of its cone. This is explained for example in Elmachtoub & Grigas (2022).

Q2 (sensitivity to κ\kappa). κ\kappa is selected on validation like the other hyperparameters. The method is not very sensitive to this choice because its extremes are known methods: as κ→∞\kappa \to \infty, the shrinkage falls back to the standard method, and as κ→0\kappa \to 0, it approaches normalization to radius κ\kappa. A poor choice of κ\kappa therefore drifts towards one of these baselines rather than causing a collapse.

Q3 (imitation and regret). We thank the reviewer for this question. Imitation is harmless when θˉ\bar\theta is a deterministic function of the input features and the model can reproduce f(θˉ)f(\bar\theta) on every sample. Otherwise, imitation discards the cost information in θˉ\bar\theta and weights decision errors by their distance in decision space rather than by their cost (Bengio et al., 2021). It thus increases regret when costs are noisy given the features, since it fits the distribution of optimal decisions rather than the decision of minimal expected cost, and when the model class cannot reproduce f(θˉ)f(\bar\theta) on every sample, since it may then accept rare but costly errors. A paragraph has been added in Section 3.2 to clarify why degeneration into imitation-based learning is not desirable.

Dear Reviewer, as the discussion period ends on November 18, we would be grateful if you could let us know whether our response addresses your comments. Following your suggestion, we added norm clipping and an explicit L 2 penalty as baselines, and fully described the benchmarks in Appendix E. We are happy to clarify any remaining point.

I have read your responses and I maintain my original (positive) view of your paper.

Reviewer_6zW7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines scale mismatch between predicted costs and perturbations in decision-focused learning (DFL). The authors argue that changes in the magnitude of predicted costs during training can make perturbation-based gradient estimates less informative. To address this, they propose a smooth shrinkage function that bounds the norm of predicted costs without changing the resulting optimal decision. The paper analyzes this issue using solution stability and scaling invariance.

The proposed regularization is evaluated with three DFL methods (SPO, DPO, and DBB) on a three-vertex problem and two standard benchmarks. While the results show improvements in some settings, particularly when the unregularized methods fail, the gains are inconsistent across benchmarks. Overall, the paper presents a simple way to control the relative scale of costs and perturbations, but the significance of the problem and the benefits over existing approaches remain unclear.

Critical strengths

S1. Identifying a potential issue in perturbation-based DFL.

The paper highlights how changes in the relative scale of predicted costs and perturbations can affect learning. The stability-ratio analysis and the three-vertex experiment help illustrate this behavior.


S2. Simple regularization.

The proposed shrinkage is straightforward to implement and can be added to existing DFL methods without changing their gradient estimators. It also preserves the original optimal decision while controlling the norm of the predicted costs.

Critical weaknesses

W1. Unclear practical importance of scale mismatch.

There is not enough evidence that scale mismatch is a serious problem in existing DFL methods. While the issue itself is reasonable, the clearest evidence comes from a three-vertex problem specifically designed to exhibit training instability. It is less clear how often this problem arises in practice or how much it affects the performance of standard DFL methods. The paper also does not convincingly explain why existing ways of controlling perturbation scales are insufficient.


W2. Lack of novelty in the proposed regularization.

The proposed method is a simple rescaling of predicted cost vectors. Cost normalization is already a common technique, and the proposed shrinkage has a similar effect to adjusting the perturbation scale relative to the predicted costs. The paper does not clearly show what makes this approach different from existing normalization methods or why it should work better than other simple ways of controlling scale. As a result, I find the methodological contribution rather weak.


W3. Weak theoretical contribution.

Much of the theoretical analysis relies on well-known properties of linear optimization. Proposition 4.1, for instance, follows directly from positive scaling invariance, and the stability ratios mainly describe the relative scales of predicted costs and perturbations. More importantly, the analysis does not establish that the proposed shrinkage leads to better gradient estimates or lower regret. Shrinkage also improves one of the proposed stability ratios while potentially making the other worse. Overall, I find that the theory offers little beyond explaining the basic effect of rescaling.


W4. Concerns about the gradient analysis.

Section 4.2, particularly the arguments around Equations (9) and (10), needs further justification. The paper uses the stability of exact optimal decisions to explain when gradients involving expected perturbed decisions vanish or become uninformative. However, a decision remaining unchanged under a particular perturbation does not necessarily imply that its expectation remains unchanged under a perturbation distribution. The conclusions about gradient collapse and degeneration into imitation therefore seem to require additional assumptions. This is concerning because these arguments play an important role in motivating the proposed method.


W5. Insufficient experimental evidence.

The experimental evaluation is rather narrow, covering a specially designed three-vertex problem and only two standard benchmarks. Although the proposed regularization improves performance in some settings, the results are mixed, and standard L2 normalization performs better in several cases. The paper also acknowledges in Section 6.2 that it cannot directly verify whether scale mismatch is responsible for the observed improvements on standard benchmarks. This makes it difficult to connect the experimental results to the main argument of the paper.

There is also an inconsistency in Section 6.2. The DPO/SP3 improvement from 125.57 to 13.10 is attributed to the proposed shrinkage, but Table 2 reports 13.10 for L2 normalization and 15.72 for shrinkage. Overall, I do not find the experiments sufficient to establish the practical value of the proposed method.

Questions for authors

Q1. Significance of scale mismatch.

How strong is the evidence that scale mismatch is a common and important cause of training failures in DFL? While the three-vertex experiment illustrates the issue, its role in the standard benchmarks remains unclear.


Q2. Theoretical justification and guarantees.

Which assumptions are needed to justify the gradient analysis in Section 4.2, particularly Equations (9) and (10)? More importantly, under what conditions does the proposed shrinkage prevent gradient collapse or improve learning? The current analysis mainly describes changes in stability ratios without establishing how these changes lead to better optimization performance.


Q3. Advantages over simpler alternatives.

Does the proposed shrinkage offer a clear advantage over existing normalization or sample-dependent perturbation scaling? A direct comparison would help establish whether it provides a meaningful improvement.

5 comments
AuthorsOpenReview ↗

Answers to Reviewer 1, part 1.

We thank the reviewer for their careful reading and constructive comments, and we propose to answer all comments below.

W1. We would like to emphasize that the mechanism follows from the theory: by scaling invariance and Lemma 4.1, the ratio between the perturbation scale and ∥θ∥\|\theta\| is set per sample, the loss does not constrain it, and it can drift during training.

The stability ratios themselves cannot be computed for larger optimization problems: this is the purpose of the three-vertex problem. However, we can measure how often the failure occurs during training, i.e. how often perturbed decisions coincide with the unperturbed one. We do so in our new experiments with the decision-change rate (Section 6.2, Tables 4–5), which establish the impact of this instability on standard DFL benchmarks with realistic optimization problems and datasets.

Finally, all experiments now consider normalizing the ground-truth cost vectors as a preprocessing step, selected on validation. Our results show that it is not sufficient to address this instability: regularizing the estimates appears necessary to keep perturbations and cost estimates on the same scale.

W2. We agree that cost normalization is a common technique, and we do not claim the rescaling operation itself as our contribution. Our contribution is to explain why regularization is necessary in perturbation-based DFL, beyond the settings where it has been used so far.

Because the optimization mapping is scale-invariant, one might expect regularizing the predicted costs to be pointless, or at best a harmless convention. We show this is not the case once additive perturbations are introduced: the norm of the predicted costs then sets the effective perturbation scale of each sample. Since the regret does not constrain this norm, the effective scale drifts across samples and over training. This is a dynamic problem that a fixed hyperparameter cannot correct. Prior uses of normalization (Sahoo et al., 2022; Schutte et al., 2026; Wan & Liu, 2026) address static concerns tied to their specific methods, not this drift. Table 1 shows that regularization is far from neutral.

We address alternative normalizations with our new experiments. First, target cost normalization is treated as a hyperparameter and selected on validation. Second, an L2L_2 loss penalty is studied as an alternative to our regularizations. Where the standard method, with or without target normalization, is ahead of the shrinkage rsr^s, the difference is small (less than one point) and never significant, whereas the shrinkage gains reach a factor of five (DPO, c3). This is expected: target normalization is only a preprocessing step and cannot handle the dynamic scale mismatch between cost estimates and perturbations. The L2L_2 penalty does steer the norm of the cost estimates during training, but it offers no guarantee of keeping it below a threshold, and it is less reliable: it is worse than the standard method for DBB on every benchmark and for SPO in c3, and it falls far behind clipping for DPO in c3.

W3. We agree that Proposition 4.1 follows directly from positive scaling invariance, and we now present it as a lemma (Lemma 4.1). Its role is to separate the stability radius into a problem-independent factor, ∥θ∥\|\theta\|, and a problem-dependent factor, η(θ^)\eta(\hat\theta), which reflects the combinatorial geometry. Because η(θ^)\eta(\hat\theta) depends on the instance and on the predicted direction, no general statement about regret can be made without strong structural assumptions (cf. Aubin-Frankowski et al., 2026). A general analysis can therefore hardly tie shrinkage to better gradient estimates or lower regret; this can however be done for particular cases, as for the three-vertex problem. For the same reason, ∥θ∥\|\theta\| is the only lever that can be controlled per sample regardless of the problem, and it is the lever our regularization acts on.

On the trade-off between the two ratios: as stated in Section 4.3, only one of them (ρη\rho_\eta) requires regularization. When ξη<1\xi_\eta < 1, the gradient does not vanish and the failure can be handled by hyperparameter choice. When ρη<1\rho_\eta < 1, the gradient vanishes or degenerates, and no static hyperparameter can prevent this, because ∥θ∥\|\theta\| drifts.

AuthorsOpenReview ↗

Answers to Reviewer 1, part 2.

W4. We thank the reviewer for this important remark. Two clarifications. First, SPO and DBB, two of the three methods in our experiments, involve no expectation: their gradients are deterministic, and the stability argument applies exactly. Second, for DPO and IMLE, the expectation is estimated in practice with a finite number of sampled perturbations. The stability argument therefore applies term by term to the gradient estimate that is actually computed: gradient collapse occurs whenever all sampled perturbations remain within the stability radius. This happens when the perturbation scale is no longer commensurate with the scale of the cost estimate. Drawing more perturbation samples may mitigate the issue, but only at the cost of additional solver calls. A remark to this effect was added at the end of Section 4.2.

W5. We thank the reviewer for spotting the inconsistency in Section 6.2. All experiments have been rerun under a single protocol (ten seeds, selection on validation, paired Wilcoxon tests), and the text now matches Table 2. To link the gains to scale mismatch on the standard benchmarks, we added the decision-change rate Δ\Delta. For standard DPO, Δ\Delta falls from 0.54 to 0.35 (shortest path) and from 0.34 to 0.20 (knapsack) over training, while with cost regularization it stays at 0.60 and 0.43. For standard DBB, Δ≈0.01\Delta \approx 0.01, i.e. its gradient vanishes for almost all samples; regularization raises it to 0.10 on shortest path (Tables 4–5).

Q1. Our new experiments show that scale mismatch arises on standard benchmarks, even with normalized targets, and grows with the heterogeneity of the cost scale. Bounding the norm of the cost fed to the perturbation works best: the shrinkage rsr^s, the safer default, or clipping rcr^c, stronger under extreme instability, improves DPO everywhere, SPO on the three-vertex problem and DBB on most benchmark settings. The normalization rnr^n fails on small cost vectors, the L2L_2 penalty lags where instability is strongest, and target normalization does not prevent the collapse.

Q2. We clarified the gradient analysis in our answer to W4. More broadly, no strong and general guarantee that regularization prevents gradient collapse or degeneration can be given: the stability radius, which drives the instability, depends entirely on the optimization problem under study. What holds in general concerns the stability ratios, as detailed in our answer to W3.

Q3. First, by scale invariance, the shrinkage and clipping are per-sample perturbation scalings: f(rs(θ)+δ)=f(θ+(1+∥θ∥/κ) δ)f(r^s(\theta)+\delta) = f\big(\theta + (1+\|\theta\|/\kappa)\,\delta\big), so the effective perturbation scale moves from a fixed value for small norms to a norm-proportional one for large norms. Second, our new experiments also consider target cost normalization and show that it is insufficient to address scale mismatch. Finally, as stated in our answer to W2, we do not claim the rescaling operation itself as our contribution: our contribution is to explain why regularization is necessary in perturbation-based DFL, beyond the settings where it has been used so far.

Dear Reviewer, as the discussion period ends on November 18, we would be grateful if you could let us know whether our response addresses your concerns. In particular, we proposed more relevant experiments, added the decision-change rate to measure scale mismatch directly on the standard benchmarks, corrected the inconsistency you identified in Section 6.2, and made the trade-off between the two stability ratios precise (answer to W3). We are happy to clarify any remaining point.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Official Comment by Authors

We thank the reviewers for their careful reading and constructive comments, which led to a substantial revision.

Our contribution in brief. Perturbation-based DFL methods share an overlooked failure mode. Because decisions are invariant to the scale of the predicted costs, nothing in the loss controls that scale, and it drifts relative to the fixed perturbation, across samples and over training. Gradients then vanish or degenerate into imitation. A global perturbation scale cannot follow a per-sample drift; bounding the norm of the cost fed to the perturbation can, through a simple, method-agnostic transformation that leaves decisions unchanged.

What the revision adds.

  • Direct evidence on standard benchmarks. we consider two new sets of PyEPO benchmark experiments on the knapsack and shortest path problem that provide further evidence of our approach relevance. A new solver-based diagnostic, the decision-change rate Δ\Delta, measures the collapse directly anf demonstrates that the proposed regularization effectively addresses instability when it arises. For example, standard DBB on knapsack has Δ≈0.01\Delta \approx 0.01: its gradient vanishes for almost every sample. For standard DPO, Δ\Delta drops by 35–40% during training. Cost regularization prevents this collapse and significantly improves performance.
  • Stronger baselines. We added norm clipping, an explicit L2L_2 penalty on predicted costs, and the possibility to apply target normalization depending on validation for every method. Target normalization does not prevent the collapse, and the L2L_2 penalty is unreliable (worse than the standard method for DBB on every benchmark). Clipping performs on par with shrinkage, confirming that what matters is bounding the cost fed to the perturbation.
  • Rigorous protocol. All experiments were rerun with ten seeds, selection on validation and paired Wilcoxon tests. The benchmarks are fully described in Appendix E.
  • Clearer scope and analysis. The abstract and introduction now position our setting, linear cost prediction for combinatorial problems, within the broader DFL literature. We also clarify how the stability argument applies to sampled gradient estimates.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.