acceptodds
Under review as a conference paper at ICLR 2027

Refined Label Comparisons for Constrained Non-Monotone -Submodular Maximization

Abstract

Non-monotone -submodular maximization under support constraints requires controlling both feasibility conflicts and label-dependent losses. We develop Random-root, a minimax randomized greedy method using support budgets, and a continuous algorithm that refines these comparisons through label-specific replacement losses. Under a matroid, the continuous algorithm improves the approximation coefficient from to , up to any accuracy loss . For every and fixed accuracy, the baseline uses value queries for elements and rank . Refinements attain for three to five labels and a uniform guarantee for all . More generally, for orthant-submodular objectives with -wise monotonicity, Random-root yields a closed-form approximation guarantee on -systems, while continuous optimization attains under matroid constraints, with for . For , the extra-label refinement saturates at : all share the same exact replacement-loss envelope.

Primary area:
optimization
Keywords:
non-monotone k-submodular maximization, continuous optimization, matroid constraints, p-systems

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_7Z9hOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies non-monotone k-submodular maximization under support constraints, and more generally full-domain orthant-submodular objectives satisfying r-wise monotonicity. Its first contribution is Random-root, a randomized greedy algorithm whose label distribution solves a local minimax problem based on support-charging losses. The algorithm obtains an explicit approximation factor R(p,r)R(p,r) for p-system constraints with O(nkρ)\mathcal{O}(nkρ) value-oracle queries, where ρ is the maximum feasible support size. Its second contribution is a matroid-constrained continuous method using label-specific replacement losses and an adaptive comparison-weight potential. For non-monotone k-submodular objectives, it improves the recent 2−1≈0.4142\sqrt 2−1 \approx 0.4142 matroid guarantee to α2−τ\alpha_2−\tau, with α2≈0.4218\alpha_2 \approx 0.4218, and obtains α2+−τ≈0.4366−τ\alpha_2^+−\tau \approx 0.4366−\tau for k=3,4,5k=3,4,5 and a guarantee above 0.4259−τ0.4259−\tau for all k≥3k\geq3. The analysis further characterizes the exact local replacement-loss envelopes: for r≥3r\geq3 the envelope saturates once k>rk>r, whereas for r=2r=2 the resulting coefficients plateau at three to five labels and then decrease. The paper supplies proofs, finite oracle implementations, and numerical checks of the theoretical coefficients, but no empirical evaluation of optimization performance on benchmark instances.

Critical strengths
  1. The paper establishes genuinely new approximation guarantees in a well-studied setting. The improvement over the 2−1\sqrt 2−1 matroid guarantee of Chen et al. (2026b) is modest for k=2k=2 but mathematically nontrivial, and the stronger coefficients for k≥3k\geq3 are more substantial. Random-root also extends the guarantee to pp-systems and to rr-wise monotone orthant-submodular objectives. These are concrete improvements over the results of Ye et al. (2025), Zhou et al. (2025), Huang et al. (2026), and Chen et al. (2026b), rather than merely an alternative analysis of the same approximation factor.

  2. The label-specific comparison framework provides a substantive technical contribution beyond the final numerical constants. In particular, the distinction between support-budgeted and replacement losses (Proposition 4.1), the weighted potential transfer (Proposition 4.2), and the exact envelope classification across label counts (Section 5 and Appendix B) give a coherent explanation of when asymmetric label choices help. The paper also addresses finite implementation, estimation errors, and rounding rather than relying solely on an idealized continuous trajectory. This combination makes the work of independent theoretical interest.

Critical weaknesses

The practical benefits of the stronger approximation guarantees remain unclear. For k=2k=2, the continuous guarantee α2−τ\alpha_2-\tau exceeds 2−1\sqrt{2}-1 only when τ<α2−(2−1)≈0.00758\tau<\alpha_2-(\sqrt{2}-1)\approx0.00758. The finite continuous algorithms have value-oracle complexity bounds with a τ−3\tau^{-3} dependence, whereas Random-root achieves O(nkρ)O(nk\rho) queries without an accuracy parameter. Although these are worst-case upper bounds and both methods have comparable asymptotic dependence on n,k,ρn,k,\rho for fixed τ\tau, the additional sampling and continuous optimization may entail substantial computational overhead. Moreover, attaining some of the stronger label-dependent guarantees requires more elaborate local optimization, including a linear program for the exact β2,k\beta_{2,k} coefficient when r=2r=2 and k≥4k\ge4. The manuscript provides no instance-level comparisons of objective quality, oracle usage, or running time against Random-root or existing methods; its figures illustrate theoretical coefficients rather than algorithmic performance. Consequently, while the theoretical improvements are valid and potentially significant, their practical benefits relative to simpler alternatives remain uncertain.

Questions for authors
  1. Could the authors clarify the computational trade-off involved in obtaining a provable improvement over the existing 2−1\sqrt{2}-1 guarantee under general matroid constraints? For k=2k=2, the guarantee α2−τ\alpha_2-\tau exceeds 2−1\sqrt{2}-1 only when τ<α2−(2−1)≈0.00758\tau<\alpha_2-(\sqrt{2}-1)\approx0.00758. Although Theorem 6.2 provides a value-oracle complexity bound, it remains unclear how the required accuracy translates into practical oracle and arithmetic costs, including sampling, preprocessing, and rounding. A limited compute-matched comparison with Random-root and Chen et al. (2026b), using non-monotone instances of objective (3) under non-uniform matroid constraints and reporting achieved objective values, oracle calls, and running times, would help assess whether the refined continuous algorithm offers meaningful benefits at realistic computational budgets. Such a comparison would directly evaluate the proposed algorithmic contribution without requiring a broader application study.

  2. Two very recent preprints, Aggarwal [1] and Wang & Wang [2], report stronger approximation guarantees for nonnegative non-monotone kk-submodular maximization under matroid constraints. Could the authors clarify how the contributions of the proposed label-specific replacement framework differ from these concurrent approaches, particularly regarding the exact envelope classification, broader rr-wise objective classes, and pp-system constraints? A brief comparison would help assess the independent significance and current positioning of the results.

[1] Geometry-Dependent Approximation for Non-Monotone k-Submodular Maximization.arXiv:2610.04049.

[2] An Analytic 0.48-Approximation for Non-Monotone k-Submodular Maximization under a Matroid Constraint. arXiv:2610.20140.

4 comments
AuthorsOpenReview ↗

Certified costs and compute-matched comparisons

Thank you for the detailed assessment and the concrete suggestion for a compute-matched comparison. We have made the precision-dependent costs more explicit and added the requested two-label experiment alongside the general rr-wise experiments. Selected substantive revisions are highlighted in blue; all locations below refer to the revised manuscript.

Cost of a provably stronger guarantee

The original presentation emphasized the O(nkρ)O(nk\rho) query order at fixed accuracy. Section 6 now explicitly connects the desired guarantee to the additional precision cost. Your two-label threshold is correct: under the theorem's sampling prescription, a strict improvement requires τ<α2−(2−1)≈0.00758.\tau < \alpha_2-(\sqrt{2}-1)\approx0.00758. Theorem 6.1 (formerly Theorem 6.2) and the discussion on p. 9 show that, for fixed rr, the sparse continuous bound adds a precision factor τ−3log⁡(1/τ)\tau^{-3}\log(1/\tau) relative to Root's O(nkρ)O(nk\rho) bound. Appendix D.6 gives complete sufficient query budgets for a target coefficient γ\gamma, using tolerances αr−γ\alpha_r-\gamma and ar,k−γa_{r,k}-\gamma for the sparse and refined full-row routes, respectively.

The accounting includes preprocessing, sampled contexts, and output evaluation. Beyond oracle calls, Appendix D.5 accounts for the O(n)O(n) construction of an explicit sampled context, while Appendix C.4 describes support exchanges and finite-bit sampling. The local LP is specific to the fixed-label refinement for r=2,k≥4r=2,k\ge4; the experiments use the sparse continuous baseline and do not invoke that LP.

The requested two-label, non-uniform-matroid experiment

Appendix F.3 (p. 46) uses a modular-coverage special case of the original objective (3), retaining nonnegativity, non-monotonicity, and k=r=2k=r=2. Its 16 elements form eight pairs with at most one selected per pair, a non-uniform partition matroid of rank eight. We test six instances under each of two audience arrangements, with three independent randomized trials per instance.

In each trial, one complete continuous run determines the time allowance TT. Random-root and Chen each start from empty and spend the same allowance on their own restarts, retaining their best feasible return. All methods use the same value oracle. Continuous optimization is charged for initialization, preprocessing, sampling, updates, compression, rounding, and output evaluation.

In the high-competition group, the median allowance is 0.756 seconds. Time-matched Root, continuous optimization, and Chen attain 95.86%, 97.59%, and 100% of OPT, with median query counts of approximately 428,000, 100,000, and 457,000, respectively. Chen is strongest in this group; under ordinary overlap all three methods reach OPT. Table 3 also reports equal-query results. The cost breakdown below it shows that continuous updates, including sampling, compression, and rounding, take a median 0.751 seconds; Greedy initialization and scale preprocessing take 0.45 ms and 0.26 ms.

Additional value in the general rr-wise setting

Section 7 examines strict r=3,6r=3,6 instances. In Figure 1's high-overlap r=6r=6 group (p. 10), continuous optimization and time-matched Root restarts reach 90.34% and 84.48% of OPT. Continuous exceeds Root's full-budget best in 17 of 18 paired trials; among those 17, the median fraction of the budget needed to first exceed it is 26%. All six instances improve on average, and the equal-query comparison in Appendix F.2 also favors continuous optimization.

The empirical runs use fixed τ=0.1\tau=0.1 and N=2N=2 samples per marginal estimate to measure performance at limited computation. The stronger worst-case guarantee uses the target-dependent sampling prescription above. These complementary analyses quantify both the sufficient cost of certifying an improvement and the better feasible solutions obtained in practice under matched budgets.

AuthorsOpenReview ↗

Concurrent work and independent contributions

Regarding the two recent preprints, Section 1.2 (pp. 2-3) now cites and compares both. Our original introduction foregrounded the r=2r=2 coefficient improvements, while not sufficiently highlighting the contributions for the general case of rr; the revised abstract and contribution section give greater prominence to the general results already present in Theorems 2.1-2.2.

General objective class and support constraints

We study full-domain rr-wise monotone orthant-submodular objectives, with kk-submodularity as the r=2r=2 case. For r>2r > 2, several labels of the same item may simultaneously have negative marginals. Random-root optimizes a local comparison incorporating both label choice and support-deletion costs. Its cumulative analysis yields the closed-form guarantee on general pp-systems in Theorem 2.1, using O(nkρ)O(nk\rho) value queries.

On the common normalized nonnegative class under a matroid, Random-root improves the existing general rr-wise guarantee from 1/(r+1)1/(r+1) to 1/(1+r)1/(1+\sqrt{r}) (Table 1, p. 2). The continuous algorithm further improves the baseline coefficient for every r≥2r\ge2, with finite implementations attaining it up to any prescribed accuracy loss.

Exact replacement structure

Proposition 4.1 quantifies the local saving from retaining the displaced label. Section 5 and Appendix B determine the exact optimized replacement envelope through both attaining actions and matching upper bounds. For r≥3r\ge3, throughout the parameter interval used by the algorithm, Kr,k(λ)=Kr,r+1(λ)(k>r).K_{r,k}(\lambda)=K_{r,r+1}(\lambda)\qquad(k > r). Thus, the first additional label supplies the full coefficient improvement available from this local replacement comparison. The pairwise case has a different structure: a three-to-five-label plateau followed by a decline. These results explain how label availability changes the optimal local action and its resulting guarantee, beyond reporting a final approximation number.

Relation to the concurrent approaches

Aggarwal (arXiv:2610.04049v2) derives coefficients depending on the geometry of the feasible support region, measured by the largest selection probability feasible uniformly for all elements. Wang et al. (arXiv:2610.20140v2) combine a top-two label rule with a nested-chain potential and maximum-weight matroid bases. Our route instead optimizes the local comparison over the general rr-wise domain and classifies its exact replacement envelopes, using cumulative support losses and a matroid potential to obtain the global guarantees.

On the common nonnegative, non-monotone kk-submodular single-matroid problem, the concurrent works report stronger coefficients, including Wang et al.'s 0.48 guarantee; the revised related work explicitly acknowledges this. Our independent contributions are the algorithms and improved guarantees for general constrained rr-wise objectives, the pp-system result, and the exact label-replacement structure. Theorem 6.1 and Appendix C.5 connect these structural results to finite algorithms with explicit query bounds.

AuthorsOpenReview ↗

Follow-up on our responses

Thank you again for your detailed suggestions. As the discussion period draws to a close, we would appreciate it if you could take a look at our two responses and the revised manuscript.

They address the computational cost of the stronger guarantee, the requested comparison with Root and Chen in Appendix F.3, and the relationship to the two recent papers.

We would be grateful to hear whether these responses address your main concerns, or whether there is a particular point we could clarify further.

Thank you for your time.

Thank you for the detailed responses and revisions. The additional complexity analysis and compute-matched experiments substantially address my concerns about computational practicality, although the two-label results do not show a clear advantage over Chen et al. (2026b). The comparison with concurrent work also clarifies the paper's independent theoretical contributions, particularly the general (r)-wise framework and exact replacement-loss envelopes.

I will revise my recommendation to Weak Accept.

Reviewer_Z1QGOpenReview ↗
Rating
3
Confidence
1(−1)
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies constrained non-monotone k-submodular maximization and extends the analysis to orthant submodular functions satisfying r-wise monotonicity. It proposes Random-root, a randomized greedy algorithm with approximation guarantees under p-system constraints, and a continuous optimization framework using label-specific replacement comparisons under matroid constraints. The latter improves the stated approximation guarantee for non-monotone k-submodular maximization and establishes refined bounds depending on the number of labels. The paper also analyzes finite implementations and oracle complexity. Validation is primarily theoretical, supplemented by numerical evaluations of approximation coefficients and illustrative examples.

Critical strengths
  1. Improved approximation guarantees with meaningful theoretical contributions. The paper improves the previously reported approximation bounds for constrained non-monotone k-submodular maximization and extends the analysis to the more general class of r-wise monotone orthant submodular functions. These improvements address a well-established theoretical optimization problem.
  2. A systematic framework connecting local comparisons to global guarantees. The label-specific replacement-loss analysis, minimax randomization, and decreasing-weight potential provide a coherent approach to algorithm design. The characterization of label-count-dependent approximation coefficients and the accompanying oracle complexity analysis further strengthen the theoretical contribution.
Critical weaknesses

None

Questions for authors
  1. Could the authors clarify how the comparison inequalities and matroid exchange arguments in Proposition 4.2 remain valid when the comparison and optimization states assign different labels?
  2. Could the authors clarify how Algorithm 2 controls accumulated errors from sampling, discretization, compression, and rounding? Numerical experiments on representative instances would also help demonstrate the practical benefits of the improved guarantees.
1 comment
AuthorsOpenReview ↗

Label comparisons and finite-implementation errors

Thank you for the positive assessment and the questions about Proposition 4.2 and Algorithm 2. We have expanded the explanations linking the comparison construction, support exchange, and finite error bounds, and added instance-level experiments. Selected clarifications and additions are highlighted in blue; references below are to the revised manuscript.

1. Different labels in the comparison and optimization states

The original manuscript provided probabilistic expressions for comparing states, but the explanation of their common structure and the auxiliary basis preserved during the sequence permutation process is rather brief. Sections 4.1-4.2 (pp. 5-7) now separate three operations: comparing the same label across compatible contexts, accounting for label replacement, and matching support elements.

In Eq. (7), fix an optimal labeling oo with support OO. The coupled comparison draw retains every active label of the algorithmic draw XxX_x and fills only inactive elements of OO with their fixed optimal labels. Deleting the updated element ee gives X−e⪯Z−eX_{-e}\preceq Z_{-e}. Applying OSM to the same candidate label jj and taking expectations therefore yields

\le \mathbb E\Delta_{e,j}F(X_{-e})=g_{e,j}.$$ The $r$-wise inequalities also survive averaging, so $a_e\in\mathcal A(g_e)$. The algorithm need not choose $o_e$. Equation (8) explicitly accounts for this difference: when $e\in O$, the comparison increment is $\delta(q_e\cdot a_e-a_{e,o_e})$; when $e\notin O$, it is $\delta q_e\cdot a_e$. The replacement bound in Eq. (9) covers both. Thus, we do not apply OSM directly between incompatible label assignments. Matroid exchange handles a different object: the auxiliary **support** base $C_t=P_t\mathbin{\dot\cup}R_t$, where $P_t$ is the direction prefix and $R_t\subseteq O$ contains unmatched optimal elements. The labeling $o$ used in Eq. (7) stays fixed. Each selected $e_t$ is matched to a distinct $u_t\in O$ that was addable at selection, giving $$m_{e_t}(x^{t-1})\ge m_{u_t}(x^{t-1})\ge m_{u_t}(x^{\mathrm{end}}).$$ The first inequality follows from greedy selection and the second from OSM along the block. Each optimal element is matched once; summing the local gains and using Eq. (15) pays for the endpoint weight change. The proof on p. 7 and Appendix C.1 (p. 30) give the construction explicitly. **2. Sampling, discretization, compression, and rounding** The original proof provided the final error decomposition, but the explanation of the relationships among the various error budgets was rather fragmented. Section 6 now makes the roles of the subroutines and their conditional error accounting explicit; Eq. (19) (p. 9) summarizes the guarantee, with the complete proof in Appendix C.6. The finite parameters satisfy Eq. (18): $$0\le\lambda_i-\lambda_{i+1}\le\delta_iK(\lambda_i), \qquad \sum_i\delta_i=1.$$ Row affinity gives exact finite row increments, and Proposition 4.2 shows that their gains pay for the endpoint weight decrease. The construction starts at $\lambda_0\ge a-\tau/2$, where $a$ is the ideal coefficient of the chosen envelope. Preprocessing bounds residual marginal ranges and the total caps over a base, denoted $U(A)$. Fresh samples are taken conditional on the current history. The blockwise bound is $$\mathbb E[\Phi_{i+1}\mid\mathcal H_i] \ge\Phi_i-\delta_iU(A)b_i.$$ The sparse and full-row prescriptions in Appendices D.5 and C.5 make the total expected local loss at most $\tau M/4$, including estimation failures and probability quantization. Label compression preserves conditional row means and hence the potential's conditional expectation; later blocks sample at the resulting state. Ideal support rounding uses randomness hidden from optimization and does not decrease expected value. The bounded-bit implementation is coupled to the exact process, costing at most another $\tau M/4$. Consequently, $$\mathbb E F(Y)\ge (a-\tau/2-\tau/4-\tau/4)M=(a-\tau)M.$$ These conditional bounds account for the algorithm's adaptive random history, rather than treating estimated marginals as exact. For the requested empirical evidence, Section 7 adds matched-time and matched-query comparisons. In Figure 1's high-overlap $r=6$ group, continuous optimization reaches 90.34% of OPT versus 84.48% for time-matched Root restarts, with improvement on average on all six instances. Appendix F.2 reports the fixed-query comparison and distinguishes the fixed experimental sampling effort ($N=2$ per marginal estimate) from the target-dependent sampling prescription used to certify the theorem.
Reviewer_edkCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies non-monotone kk-submodular maximization under support constraints and extends the analysis to orthant-submodular objectives. It proposes Random-root for pp-system constraints and a continuous algorithm using refined label-specific replacement-loss comparisons under matroid constraints. The algorithms improve existing approximation guarantees, achieving approximately 0.4218−τ0.4218-\tau in the baseline case and over 0.4366−τ0.4366-\tau for three to five labels, with explicit query-complexity bounds.

Critical strengths
  • The paper provides theoretical improvements over existing approximation guarantees and extends the analysis to broader constraint and objective classes.

  • The label-specific replacement-loss analysis is technically interesting and leads to stronger approximation coefficients.

Critical weaknesses
  • The novelty relative to closely related work could be clarified. In particular, the paper should more explicitly distinguish its technical contributions from prior label-specific comparisons and continuous-greedy methods.

  • The evaluation is limited to theoretical analysis and numerical verification of approximation guarantees. The paper also lacks a clear motivation for why this problem is important and what practical applications it may serve. While the problem may be primarily theoretical, the authors should explain its practical relevance and discuss whether meaningful empirical evaluation is feasible. If so, application-specific experiments or examples would help demonstrate the value of the proposed improvements.

Questions for authors

Please check the weakness section and respond accordingly.

2 comments
AuthorsOpenReview ↗

Contributions, motivation, and empirical evidence

Thank you for your careful reading and for asking us to clarify the technical contribution and practical motivation. We have reorganized the introduction and related work and added campaign experiments. Selected substantive revisions and clarifications are highlighted in blue; all locations below refer to the revised manuscript.

1. Technical contribution and relation to prior methods

The original presentation led with the r=2r=2 coefficient improvements, giving less prominence to the general rr-wise results already established in Theorems 2.1-2.2. Sections 1.1-1.2 (pp. 1-3) now organize the contribution around constrained rr-wise monotone orthant-submodular (OSM) maximization, with kk-submodularity as the r=2r=2 special case.

Random-root jointly accounts for support conflicts and label choice. Theorem 3.1 (p. 4) exactly minimizes worst-case local comparison loss per unit of expected gain, determining its adaptive uniform-prefix label rule. The cumulative loss analysis in Theorem 2.1 then yields the closed-form pp-system guarantee. On the common class of normalized nonnegative objectives under a matroid, this improves Huang et al.'s 1/(r+1)1/(r+1) guarantee to 1/(1+r)1/(1+\sqrt{r}), changing the dependence on rr from order 1/r1/r to order 1/r1/\sqrt{r}, while retaining scan-based Greedy's O(nkρ)O(nk\rho) value-query bound (Table 1, p. 2).

The continuous algorithm further exploits information about the displaced label. Instead of using the largest marginal m(g)m(g) to bound this loss, it retains the displaced label's marginal aja_j. Proposition 4.1 and Eq. (12) (p. 6) quantify the exact saving. This sharper comparison determines the label distribution by optimizing h−λHEh-\lambda H_{\mathrm E} at the current comparison weight; Proposition 4.2 then converts local progress into a global guarantee. The resulting baseline coefficient exceeds Random-root's for every r≥2r\ge2. Section 5 and Appendix B also establish the exact label-count structure: for r≥3r\ge3, the replacement envelope improves from k=rk=r to k=r+1k=r+1 and then saturates.

Thus, existing label comparisons, multilinear extensions, and matroid exchange supply important tools; our contribution is the optimized local randomization for the general rr-wise domain, its support-constrained guarantee, and the exact replacement-envelope classification that drives the continuous improvement. The revised Sections 1.2 and 4 make these distinctions explicit.

2. Motivation and instance-level evaluation

The original allocation example illustrated negative label marginals but remained pairwise monotone. The revised introduction (p. 1) explains the additional situations covered by r>2r > 2: a firm selects promoters and assigns promotion modes; overlapping customers create diminishing returns, while coordination costs can make several modes unprofitable. Marginals (2,−1,−1)(2,-1,-1) violate pairwise monotonicity but satisfy the three-label sum condition.

The new Section 7 and Appendix F use this same setting for controlled synthetic experiments with strict orders r=3r=3 and r=6r=6. Within each order, we vary audience overlap while fixing individual returns, recruitment quotas, and coordination costs. Exact optima provide the quality reference. All methods use the same value oracle; Random-root and Chen can restart throughout the time taken by one complete continuous run and retain their best feasible solutions.

Figure 1 (p. 10) shows where the additional computation is useful. At high overlap and r=6r=6, one Root run improves mean return from Greedy's 77.50% to 80.22% of OPT at the same query count. Allowing Root repeated restarts, continuous optimization reaches 90.34% of OPT versus 84.48% within the same time budget. All six high-overlap instances improve on average. As overlap increases, the mean relative gain over time-matched Root restarts rises from 2.26% to 6.94%. Appendix F.2 also reports gains under equal value-query budgets.

These experiments connect the theoretical improvement to competition for shared customers: label randomization improves low-cost decisions, and further continuous computation can yield better feasible solutions under matched budgets. The revised anonymous supplement includes the instance generators, verification code, and trial records.

AuthorsOpenReview ↗

Follow-up on our response

Thank you again for your feedback. As the discussion period draws to a close, we would appreciate it if you could take a look at our response and the revised manuscript.

Sections 1.1–1.2 clarify the general r-wise contributions and their relationship to prior work, while Section 7 adds campaign experiments addressing the practical motivation and empirical performance.

We would be grateful to hear whether these clarifications and results address your main concerns, or whether there is anything we could clarify further.

Thank you for your time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.