acceptodds
Desk rejected at ICLR 2027

Structure-Aware Primal–Dual and Large-Neighborhood Search in an LLM-Configured Optimization Agent

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

We study large-scale promotion optimization across multiple markets on Alibaba's e-commerce platform, where high-quality feasible policies must be generated within a short operational time window. Subsidies affect demand exponentially, tiered coupon rules induce price discontinuities, and decisions are coupled through complex constraints involving budgets, subsidies, cross-market consistency, and priority orderings among merchandise groups. We develop a portfolio of two complementary structure-aware methods. A primal–dual (PD) method relaxes shared constraints, yielding independent multi-market subproblems for individual products, and recombines their solutions to recover a feasible policy. An adaptive large neighborhood search (LNS) method jointly updates products linked by tight budget or group-priority constraints, enabling coordinated moves across coupon thresholds that can be difficult to achieve through multiplier updates alone. We further embed the portfolio in an optimization-agent framework, where an LLM selects hyperparameters from finite predeclared sets based on a numerical summary of each instance, after which both PD and LNS are executed and the better feasible solution is returned. Theoretically, we establish exact product-wise separability of the primal–dual relaxation and develop a branch-and-bound algorithm that globally solves each product-level subproblem with an optimality certificate. On a 72-instance test suite spanning six structural families and problem sizes from 100 to products, our portfolio returns a feasible policy for every instance under a 900-second time budget per method, whereas Gurobi finds an incumbent for only 33 instances. On these 33 common-feasible instances, the portfolio's mean objective shortfall relative to Gurobi is only 0.0094%. The scalability advantage becomes more pronounced on larger instances: for the 36 instances with , the portfolio returns a feasible policy on all 36, whereas Gurobi finds an incumbent on only 6; for the 18 instances with , the corresponding counts are 18 and 1, respectively.

Primary area:
optimization
Keywords:
Large-Scale Promotion Optimization, Primal-Dual Methods, Structure-Aware Large Neighborhood Search, LLM-Configured Optimization Agent, Nonlinear Constrained Optimization, E-Commerce Promotion

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_j7jFOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies large-scale multi-market promotion optimization. The formulation includes nonlinear demand, discontinuous coupon tiers, budget and subsidy constraints, cross-market consistency, etc. The authors propose a primal-dual method, an adaptive large-neighborhood search method, and an optional LLM-based hyperparameter configurator. The authors evaluate their work via synthetic offline experiments.

Critical strengths

I think the industrial problem is potentially interesting, and the focus on finding feasible solutions under a strict time budget is practically relevant.

Critical weaknesses

I found the paper very difficult to read. A lot of the used notation and terminology is not explained or explained insufficiently, the main algorithms are absent from the main text, and the theoretical novelty is unclear. The paper is not well-positioned against the existing literature. Given the lack of strong theoretical results, the empirical evaluation looks relatively weak to me as well: the authors rely entirely on synthetic data, provide a narrow set of baselines.

Questions for authors

A lot of notation and terminology is introduced without adequate definitions. For example, in Section 2 the authors say that R_i is an “active market”, what exactly “active” means? How is it different from a regular (or non-active) market? Coupon “tiers” are never formally defined; the “demand scale” is not clearly explained (Section 2), how is it different from the standard demand? The definitions of Xi, X, and S also require the reader to reconstruct the feasible set from several later paragraphs; the terms such as “global group-order edge” are introduced without motivation or any explanation.

More importantly, the paper does not seem to be self-contained. Specifically, it repeatedly refers to Algorithms 1–3, but the actual pseudocode is not present in the main text of the paper. Since the primal-dual procedure and LNS method are the main contributions, concise versions of these algorithms should appear in the main text. A notation table and a small running example would also substantially improve readability.

I feel a lot of text of the paper is AI-generated and might have never been proof-read. For instance, even in the Abstract the authors say “Gurobi finds an incumbent on 33.” What is “33” here? Also, the main text of the paper includes the names of the programming variables that the authors used in the code implementation of their approach (such as the term “CODE” in Section 2). Those are never used later in the paper and do not improve clarity but may rather obstruct understanding of the paper.

The paper does not clearly distinguish standard consequences of Lagrangian relaxation from new results. Theorem 1 states that the Lagrangian separates product-wise after the aggregate constraints are dualized, which appears to follow directly from the formulation. Similarly, Theorem 3 uses the standard fact that a convex function over a box attains its maximum at a corner. I’d suggest to the authors to explain precisely what is novel and how it relates to existing methods.

Importantly, I think the related work section is highly insufficient, and it is not clear how the paper contributes to what is already known. The authors should contrast their method with the existing budget optimization methods such as Buchbinder et al., “Online Primal-Dual Algorithms for Maximizing Ad-Auctions Revenue”, Mehta et. al., “AdWords and Generalized Online Matching”, Feldman et. al, “Online Stochastic Packing Applied to Display Ad Allocation”, etc. There are also a large number of works considering incentives in budget management systems which I would expect the authors to address as well, e.g., Balseiro et. al., “Repeated Auctions with Budgets in Ad Exchanges: Approximations and Design” etc. For each claimed contribution, the authors should explain what prior methods already provide, why they are insufficient, and which exact technical component is new. Without this it is difficult to understand what the main contributions of the author’s paper are.

I also think that the evaluation section can be significantly improved. To start with, all reported instances seem to be synthetic. The main benchmark contains 72 instances, generally only three instances per family and problem size which feels like a toy problem for modern e-commerce platforms. Also, I could not find an evaluation on public data, or actual promotion campaigns. The authors should explain how the generator was calibrated, which real-data statistics it reproduces, and what “held-out” means for manually designed synthetic families. The baseline comparison is also narrow. Gurobi is the primary external baseline, while most other comparisons are small internal ablations. I couldn’t find any comparisons with other existing methods (see my comments about the related work above).

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

The paper studies multi-market promotion optimization with exponential demand, discontinuous coupon tiers, resource budgets, weighted subsidy bands, product-level cross-market range constraints, and merchandise-group orderings. Separator lifting permits product-wise Lagrangian decomposition. A scalable PD method combines approximate product optimization with extensive primal recovery; a certified variant uses anchor branch-and-bound and separator-corner enumeration by component. A complementary adaptive LNS+SA method coordinates complete product assignments. An optional LLM selects finite-set hyperparameters before execution. On 72 synthetic instances with up to 100,000 products, PD is feasible on 71, every LNS seed on all 72, and the monolithic Gurobi baseline on 33 within 900 seconds per run.

Critical strengths
  1. All 216 LNS runs are verified feasible. Therefore, the 72/72 coverage is not explained merely by best-of-three selection. At N>=10,000, the reported portfolio feasibility is 36/36 versus 6/36 for Gurobi. Common re-evaluation and preservation of a verified incumbent make this a credible computational result within the reported benchmark.

  2. Effective coordination with a conditional certification path. Product decomposition, coupon-regime treatment, range anchors, and component factorization form a useful solver design. The recovery and neighborhood ablations show that problem structure matters: generic PD recovery fails on the group-order ablation instances, and full LNS improves on random neighborhoods. The six small certification instances provide additional evidence for the bound-producing variant, although they do not certify the large-scale runs.

Critical weaknesses
  1. The contribution combines familiar decomposition, auxiliary separators, anchors, corner enumeration, primal repair, and adaptive LNS for one detailed application. Such a combination can be valuable, but the manuscript needs a sharper account of the new reusable principle and a stronger comparison with related promotion-specific decomposition approaches. The optional LLM component does not close this gap: the paper appropriately presents it as a workflow case study, and the selected configuration gives lower final objectives than the default for both solvers. I do not treat that outcome as a correctness failure or demand an LLM improvement the authors do not claim. It does mean that the LLM interface supplies no demonstrated additional optimization contribution.

  2. The synthetic families reflect business constraints, but quantitative calibration of demand responses and generator distributions to deployment data is not established in the manuscript. The portfolio's best objective uses one PD run and three LNS runs against one Gurobi run; each has its own 900-second, eight-core allowance. Individual results are valuable, but portfolio quality should also be compared under equal aggregate resources. The 0.0059% shortfall concerns Gurobi incumbents on common-feasible cases, not the optimum, while the certified PD results use only N=12 and a 3600-second limit. These qualifications leave the claimed operating-window value and generality insufficiently demonstrated for the breadth of the paper's positioning.

Questions for authors
  1. What is the principal reusable algorithmic advance beyond this model-specific combination, and how does it compare with prior promotion-oriented Lagrangian or GP/IP decomposition? Which aspects of the synthetic demand and constraint distributions are quantitatively grounded in the collaboration data?
  2. Under equal total core-seconds or a fixed machine-wide wall-clock budget, how do single PD, single LNS, the portfolio, and Gurobi multi-start compare? Please separate configuration, optimization, and final-verification time. Existing per-seed feasibility results should remain visible.
  3. How are one-sided coupon boundaries encoded in the Gurobi baseline relative to the common evaluator's numerical rules, and what interpretation is intended for solver bounds near a discontinuity? Which runs and formula produce the approximately 27% large-N median bound gap? These are clarification requests, not an allegation that the baseline is invalid or that its boundary handling explains the performance difference.
Reviewer_YYc1OpenReview ↗
Rating
1
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

In this work, the authors aim to solve the problem of maximizing GMV for Alibaba through e-commerce ads shown across different marketplaces, each having potentially conflicting constraints under a single budget and campaign setup. The authors propose combining a primal-dual solution with a customized version of Large Neighborhood Search (LNS) guided by an LLM agent, which either selects default hyperparameters or adjusts one of five exposed parameters: the primal-dual step size, the time reserved for final feasibility repair, the number of candidate assignments per product in LNS, the maximum group-neighborhood size, and the time limit per restricted subproblem.

The manuscript has significant writing issues, as the motivation is somewhat unclear, the problem setup is difficult to follow, and both require substantial assumptions to be made even for readers with extensive experience in e-commerce advertising. Furthermore, the primary role of the LLM and its specific contributions are not well specified or measured — it is unclear whether the LLM is simply following a predetermined heuristic recipe to remove a human from the loop, or conducting hyperparameter exploration that would be infeasible otherwise.

While the experiments are generally well set up, the main Gurobi baseline is not thoroughly explained, and the mixed empirical results are difficult to interpret. Additionally, an MILP method is mentioned in the related work, which could have served as another strong baseline.

Critical strengths

The technical novelty of the paper is largely centered around the threshold-lifting trick for each group-ordering rule and the certified product-subproblem solver — areas where the problem is mathematically well designed.

Critical weaknesses

Overall, the paper has notable readability challenges:

1\ The promotion planning problem is not thoroughly explained, requiring the reader to make several assumptions to follow the formulation. An example walkthrough would be much appreciated for ease of understanding.

2\ The practical urgency and motivation of the problem are also somewhat unclear.

3\ It is difficult to follow which constraints are ultimately incorporated into the acquisition function, as the authors leave multiple options open in Equations 4–6.

4\ The role of the LLM appears somewhat limited, and its isolated impact remains unclear.

5\ The experimental setup could be explained in greater detail; specifically, the Gurobi baseline should be better introduced, and a comparison with the MILP approach would also be beneficial.

6\ The experimental results do not clearly favor any single method.

Questions for authors

Please see critical weaknesses above.

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

Related papers

Discussion (0)

Sign in to comment.