acceptodds
Under review as a conference paper at ICLR 2027

EvoOpt-LLM: Evolving Industrial Optimization Models With Large Language Models

Abstract

Supply-chain production and transportation models must account for changing production, workforce, inventory, and delivery rules while retaining the relationships among existing decisions. We present EvoOpt-LLM, an LLM-based approach to constructing an initial optimization model, injecting additional business constraints, and predicting candidate zero fixings before numerical solution. Three independently trained LoRA adapters generate formulations and code from natural-language problems, updated models from LP files and business rules, and variable-fixing lists from the current LP text. The supply-chain application centers on linking new rules to existing production decisions and conditioning fixing predictions on the model being solved. We evaluate the modules separately. A multicategory construction corpus measures basic modeling ability: with 3,000 examples, generation reaches 91%, while executability and reference-result matching among generated models reach 65.9% and 26.37%. A workforce-update case shows how adaptation changes the organization of a generated business-rule extension. With 400 pruning examples, variable-label F1 peaks near 0.56 on models of around 350–400 lines and declines on longer inputs. The results characterize LLM assistance for initial formulation, business-rule updates, and candidate variable fixing in supply-chain modeling.

Primary area:
optimization
Keywords:
optimization modeling, large language models, constraint injection, mixed-integer linear programming

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_v5G3OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper fined tuned three open LLMs to perform three tasks in optimisation modeling: 1) constructing mathematical formulations and solver code from natural language description; 2) extending existing math models in the form of LP files with business rules; and 3) performing variables pruning. They also perform experiments to show the effectiveness of their fine-tuned models.

Critical strengths
  1. The problem they tackle may be important in industry;
Critical weaknesses
  1. Limitations on presentation and organisation:

    1. The presentation is poor. Even after reading the abstract and introduction, I did NOT know too much about this paper (What are existing things? What are the research gaps? what are their new contributions? Why should we do this rather than that? things like this are not clear...the authors seem not familiar with writing a ML paper for a top venue like ICLR).
    2. The motivation of this paper is not clear: why must we inject constraints or perform variable pruning using these fine-tuned models? How about using a capable LLM such as GPT-6.1 to do it? (What are the gains/benefits of fine-tuned models compared to direct prompt LLMs?) Besides, the authors spent so little room on explaining why we need to do constraint injection or variable pruning (rather than building a model from the scratch)? What if the efforts/costs of injecting constraints and/or pruning variables are greater than building from the scratch?
    3. The categories in Table 2 contains both mathematical-programming classes and application families. This is very poorly organised.
    4. Figure 2 and Figure 3 are of low resolution, which is unprofessional.
  2. Limitations on methodology:

    1. Their method can only tackle MILP problems (at least from the description, I feel so) while other types of optimisation problems are not discussed;
  3. Limitations on experiments:

    1. It is very confusing why the authors selected to use openPangu-Embedded-7B-V1.1 (few people use this in their research, even researchers from Huawei...) as the base model? (Why not more capable open LLMs such as Qwen-3, Qwen3.5?)
    2. They did not have any baselines (except the untuned Pangu-Embedded-7B-V1.1). This is not what a scientific paper should do.

I have the feeling that this is a student project or an unfinished intern project in a company.

Questions for authors

Please check the weak points.

Reviewer_r9a3OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EvoOpt-LLM, an LLM-based framework for supply-chain production and transportation optimization. The framework uses three independently trained LoRA adapters for three tasks:

  • Model construction: Generating optimization formulations and implementation code from natural-language descriptions.

  • Constraint injection: Updating existing LP models with additional business rules.

  • Variable-fixing prediction: Predicting candidate zero fixings from the current LP model.

The authors evaluate the three modules separately, examining construction performance across training-set sizes, constraint injection through a workforce-update example, and variable-fixing prediction across different LP input lengths.

Critical strengths
  1. Clear organization of the three modules

The paper clearly presents the three main modules. In particular, Table 1 provides a useful overview of the inputs, outputs, and subsequent uses of each module, making their roles easy to understand.

  1. Separate evaluation of the three modules

I appreciate that the authors evaluate the three modules separately. The experiments examine construction performance across different training-set sizes, compare constraint injection before and after task-specific adaptation, and analyze variable-fixing prediction across different LP input lengths.

Critical weaknesses
  1. Limited comparisons with existing methods

Although the paper discusses several existing optimization modeling approaches in the related work, it does not provide direct quantitative comparisons with these methods. Most experimental results focus on comparisons with the untuned base model or the performance of task-specific adaptation.

Comparisons with relevant existing methods would help readers better understand the advantages of EvoOpt-LLM.

  1. Limited evaluation of the overall framework

Although the paper presents three modules as part of an optimization workflow, the experiments evaluate them separately. It remains unclear how the three modules perform when used together.

An end-to-end evaluation, including an ablation study that removes each module individually, would help clarify the contribution of each module to the overall framework.

Questions for authors
  1. In Section 5.1, the pruning module uses 400 training examples, whereas the construction and constraint injection modules use 3,000 examples. Could the authors explain why the pruning dataset was constructed at a much smaller scale and discuss how this difference in dataset size might affect performance?

  2. Could the authors provide a concrete supply-chain example in Figure 1 to better illustrate how the three modules interact within the proposed workflow?

  3. Figures 2, 3, 4, and 5 appear to have low resolution. Could the authors improve the image quality to make the figures easier to read?

Reviewer_J1NXOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

EvoOpt-LLM addresses optimization-model construction and maintenance under changing supply-chain requirements. It independently trains three LoRA adapters on a shared 77B language model to generate formulations and code from natural-language descriptions, append business constraints to existing LP models, and predict candidate zero fixings from the current model. Separate modeling experiments, a workforce-update case study, and variable-label prediction experiments demonstrate preliminary potential for task-specific fine-tuning in industrial optimization, while end-to-end reliability and actual solver acceleration remain unverified.

Critical strengths

The paper provides a clear problem boundaries in equations, and the traceable reporting is very convenient for me to read. The authors also report a reasonable implementation detail, including base model, hardware, LoRA parameters, and training configuration. Appendix A provides mathematical templates for business rules and linking constraints.

Critical weaknesses

From the perspective of contribution and novelty: The method primarily fine-tunes three separate LoRA adapters on a shared base model. It does not establish a clear methodological advance over existing optimization modeling, model editing, and learning-assisted variable fixing, or demonstrate capabilities gained by combining the modules beyond using them independently. The paper addresses industrial model maintenance under changing business requirements, but supports only appending restrictions and evaluates the modules separately. A complete workflow under successive requirement changes is not demonstrated, leaving its overall value for real model maintenance unclear.

From the perspective of the experiments: The main baseline is the unfine-tuned base model, which scores zero on all three reported metrics. Reasonable few-shot configurations and comparisons with relevant methods such as OptMATH and OR-R1 under shared data and budgets are missing, making it difficult to separate formatting adaptation from improved modeling ability. Although central to the paper, injection is assessed mainly through one qualitative workforce example. Rule satisfaction and preservation of existing logic are not quantified across the five business-rule categories, leaving reliable adaptation to business changes unestablished.

Questions for authors

Some minor concerns:

  • Permissive correctness scoring: The match rate over all inputs is approximately 24.0%24.0\%. Equation (23) allows a 5%5\% deviation after rounding objective values, and the evaluation maps different non-OPTIMAL statuses to No Best Solution, potentially masking numerical errors and differences in solver status.
  • Unclear sample counts: The 3,0003{,}000 injection records in Table 6 contain only 1,3351{,}335 unique input-output pairs. Of 413413 pruning records, 400400 are used for training, but Figure 3 does not specify the actual test size or counts in each input-length range.
  • Figures hinder verification: The smoothed three-dimensional surface in Figure 3 obscures raw observations and lacks uncertainty estimates. Small text and omissions in the code excerpts in Figures 4 and 5 make links between business rules and existing variables difficult to verify. Content absent from an excerpt should not be treated as an error in the complete program.
  • Could you provide complete generated LPs for all five rule categories on an independent test set, along with parsing success, rule satisfaction, preservation of existing constraints, and the solver result corresponding to Figure 5?
  • How many test instances underlie Figure 3? If they come from the same 413413 records, how do the instances remaining after the 400400 training records support the full input-length curve? Please specify structural deduplication, date-based splits, and sample counts in each length range.
  • After jointly fixing all predicted variables, how many instances become infeasible or lose every optimal solution? How do end-to-end time and solution quality compare with the default solver, presolve, and learning-assisted fixing methods?
  • Why does the unfine-tuned base model score zero even on complete code generation? What is the net benefit of LoRA under a shared output template, few-shot COPT examples, and stronger baselines?
  • What is the modeling match rate when evaluated without objective rounding, with strict numerical tolerances and distinct solver-status labels?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.