acceptodds
Under review as a conference paper at ICLR 2027

Requirement-Aligned LLM Modeling for Hot Rolling Scheduling in Steel Manufacturing

Abstract

Steel manufacturing's high energy use and carbon emissions make effective production scheduling—particularly hot rolling scheduling—important for efficiency and sustainability. However, complex, dynamic, and highly specific requirements of steel production limit the flexibility of conventional optimization methods, leaving scheduling dependent on local human expertise. Large language models (LLMs) can construct and adapt models for the hot rolling scheduling problem (HRSP) from natural-language requirements, but incomplete or inconsistent formulations may be infeasible or yield schedules that violate the target production constraints. To address these limitations, we propose a requirement-aligned LLM-based framework for constructing and solving multi-objective HRSP models from natural-language descriptions and instance data. A knowledge base organizes expert formulations into reusable decision, constraint, and objective components with production semantics and dependency records. Component Selection matches components to target requirements through semantic comparison, evidence verification, and coverage checks that recover omissions. Component Composition aligns decision definitions and dependencies to assemble the selected components into a consistent constraint system and objective vector. Model and Solution Verification checks component dependencies and data references before solving, then independently validates target feasibility, objective values, and Pareto non-dominance. The framework achieves F1 scores of 0.952 for component selection and 0.966 for adaptation, with all returned schedules passing independent checks of the target constraints. Our code is available at https://anonymous.4open.science/r/llm4hrsp

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Automated optimization modeling, Multi-objective scheduling, Large language models

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_ejxvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed a LLM-based modeling system, specifically for hot rolling scheduling for steel manufacturing. A large knowledge base is pre-built about how different types (components) of variables, constraints and objectives would look like, and the connection between them (e.g., to use a specific objective component, certain variables must be created in advance). When prompted to generate an optimization model, the system use RAG to fetch candidate components from the knowledge base, and constructed the model based on thses components.

Critical strengths
  • This paper addressed an important optimization problem in real world
  • The performance on the proposed scenario (hot rolling scheduling for steel manufacturing) is impressive
  • This paper suggests that a well-crafted, structured knowledge base can significantly improve the performance of LLM OR modeling
    • While the paper is very narrowly scoped on the hot rolling scheduling problem, I guess it also applies on other type of industrial OR scenarios.
  • Sufficient details

Generally speaking, I would praise it as a solid white paper, illustrating a domain-specific enterprise practice on AI for OR.

Critical weaknesses
  • The proposed approach heavily relies on a well-crafted knowledge base with sufficient expert knowledge. I guess this is the reason why this paper is narrowly scoped on the hot rolling scheduling problem for steel manufacturing.
    • This paper would be much more impressive if the proposed approach can be paired with a (semi-)automatical knowledge base generation system, and experiment on standard AI for OR benchmarks covering more scenarios.
  • Most of the techniques used in the paper, like RAG, structured KB, component composition and solution verification, are not new. This is less problematic or even encouraged in enterprise practice, but could be a very critical problem as a research paper, especially in venues like ICLR.
    • I suggest submitting to more application-oriented venues if the authors do not plan for a major revision.
Questions for authors

See "weaknesses"

1 comment

Thank you for recognizing the practical relevance and performance of our work.

[W1] [W2] Automated knowledge-base construction

The agent extracts formulas with explanations and source locations from the 20 training studies, groups them into decision, constraint, and objective components, and records production meanings and source-supported connections. Programmatic modeling validation verifies formula ownership and source references. Component grouping and connection recording are automatic; human input concerns source preparation, initial organizational guidance, and ambiguous production meanings.

[W1] [W2] Applicability beyond HRSP

We agree that broader evaluation is important and appreciate this recommendation. Component Selection and Component Composition are designed for industrial OR scheduling, with HRSP as the primary evaluated application. Transfer requires domain-specific formulations, production semantics, and data mappings. We will evaluate kettle and filtration scheduling using new component libraries and the same modeling mechanisms.

[W3] [W4] Challenges in industrial scheduling OR

The challenge is preserving production meaning as requirements change. Industrial scheduling combines context-dependent production rules with decisions coupled across stages. Relevant formulas can still introduce inappropriate restrictions, omit requirements, or leave decisions incorrectly connected. Such errors change the intended production problem even when the generated model remains solver-feasible.

Our contribution is to identify and correct these errors during model construction. Bidirectional selection checks component applicability and requirement coverage: forward matching assesses candidates against target requirements, while reverse review recovers uncovered requirements and incorrectly rejected components. Executable composition traces dependencies through auxiliary definitions and uses solver-generated counterexamples to identify declared stage relations not enforced by the assembled constraints. Missing connections trigger composition correction; missing decision dependencies trigger selection review.

The results demonstrate more precise responses to production changes. DeepSeek + RAG matches our change recall of 0.933, but its change precision is 0.250 versus 1.000 for ours: it recovers required updates while also introducing unsupported changes. Across the expanded comparison, including ORThought, MIRROR, OptiMUS, CoE, and AlphaOPT, change F1 is at most 0.394, compared with 0.966 for our method. These results support the value of explicitly maintaining requirement–component correspondence beyond retrieving relevant knowledge.

Reviewer_5JeTOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a requirement-aligned LLM framework for multi-objective hot rolling scheduling. It organizes expert formulations into reusable decision, constraint, and objective components, using component selection, dependency-aware composition, and verification to construct models and validate schedules. Experiments on five held-out HRSP cases and 15 requirement changes achieve component selection and adaptation F1 scores of 0.952 and 0.966, respectively. All returned schedules satisfy independently checked target constraints, with additional evaluations of Pareto quality and operational performance.

Critical strengths

The paper addresses a practical challenge in automating HRSP modeling under varying production requirements. Its evaluation goes beyond solver feasibility by independently checking whether generated schedules satisfy the target constraints.

Critical weaknesses
  1. The framework primarily combines knowledge-base retrieval, semantic matching, component composition, and verification, but its distinct methodological contributions remain unclear. This makes it difficult to assess its novelty beyond a domain-specific modeling workflow.
  2. The evaluation lacks comparisons with closely related agent-based optimization modeling systems and ablation studies of the proposed mechanisms. This limits the evidence for the framework's advantages over existing approaches and the contributions of its individual designs.
Questions for authors
  1. What are the key methodological contributions of the proposed framework, particularly in its component-based modeling approach, compared with existing LLM-based optimization modeling methods?
  2. How was the knowledge base constructed from the 20 training studies? To what extent were element extraction, component formation, and connection recording automated, and how much expert involvement was required?
  3. Does the framework explicitly support incremental modification of an existing model, or does it reconstruct the model from modified requirements? How are unaffected components preserved?
  4. Why were closely related agent-based optimization modeling approaches, such as OptiMUS, Chain-of-Experts (CoE), and AlphaOPT, not included in the comparisons?
  5. Given the evaluation on only five held-out HRSP studies and 15 single-requirement changes, how robust are the reported results, and to what extent might the framework generalize to other optimization problems?
2 comments

Thank you for your constructive questions.

[W1] [Q1] Primarily contribution

Our main methodological contribution is to move beyond generating executable mathematical formulations toward systematically maintaining fidelity between operational requirements and optimization models.

Specifically, we address three important modeling failures: selecting mathematically relevant but operationally inappropriate components, overlooking necessary requirements, and assembling individually valid components without enforcing their required relationships.

Our method addresses accurate model adaptation under changing production requirements. Industrial scheduling combines context-dependent rules with tightly coupled decisions. Bidirectional selection identifies components that do not match the requirements and recovers missing ones; composition uses dependency tracing and solver counterexamples to identify missing connections.

MethodInitial F1Change F1
DeepSeek + RAG0.9080.394
ORThought0.8680.255
MIRROR0.8340.286
OptiMUS0.8620.373
CoE0.8760.286
AlphaOPT0.8790.320
Ours0.9520.966

The results show a change F1 of 0.966, compared with at most 0.394 for the baselines. This measures the accuracy of requirement-induced component updates, demonstrating that our framework more accurately translates production changes into the necessary mathematical modifications.

[Q2] Knowledge-base design

We design the component-based knowledge structure to support precise correspondence between production requirements and mathematical models. Industrial scheduling involves coupled decisions and context-dependent production rules. Elements retain formulas and their production meanings; components group formulas implementing a complete decision, rule, or objective; connections capture their mathematical dependencies. This structure supports requirement-specific selection and consistent composition. The agent populates the library automatically from the 20 training studies by extracting formulas and explanations, grouping components, and recording source-supported connections.

[Q3] Model reconstruction

The implementation constructs a new model from the modified description by rerunning Component Selection and Component Composition. Newly added requirements enter matching and coverage review, and composition resolves their decision dependencies. Comparing the original and regenerated component sets identifies additions and removals. The 15 requirement-change tests include additions, removals, and modifications.

Unchanged requirements are matched against the same library using their production meanings and applicability conditions. Their corresponding components therefore remain applicable, while Stability measures whether initially correct, unaffected components are retained. Section 4.2 and Table 1 report update accuracy and Stability; Appendix C.3 provides the metric definitions.

[W2] [Q4] Comparisons with agent-based methods

We have added OptiMUS, CoE, and AlphaOPT, together with ORThought and MIRROR. The expanded comparison is presented in the table in our response to [Q1]. In the tests, the added methods’ change precision is at most 0.250, compared with 1.000 for ours.

Industrial production frequently requires adjustments due to urgent orders, temporary equipment unavailability, or revised delivery priorities. Accurately translating these changes into mathematical models is therefore essential to accommodate new requirements while preserving unaffected production rules.

[W2] Ablation study

We agree that individual contributions require controlled comparisons and appreciate this request. The additional ablation study results are summarized below:

ConfigurationInitial PInitial RInitial F1Change PChange RChange F1Stability
Full0.9940.9200.9521.0000.9330.9661.000
DeepSeek + RAG0.9140.9070.9080.2500.9330.3940.917
w/o coverage recovery0.7400.4180.5320.1780.5330.2670.924
w/o decision-dependency feedback0.7360.4080.5220.1950.5330.2860.941

Within these results, both removal variants miss required updates and introduce unsupported changes, despite retaining most initially correct, unaffected components. Their stability of 0.924 and 0.941 therefore does not establish accurate adaptation.

[Q5] Reproducibility and transfer beyond HRSP

Thank you for this suggestion. We will evaluate transfer to kettle and filtration scheduling by constructing domain-specific component libraries while retaining the same Component Selection and Component Composition mechanisms.

Thank you for the detailed response and additional experiments. The expanded comparisons with related optimization modeling methods and the new ablation studies provide stronger empirical support for the proposed framework. However, I still have concerns about the methodological novelty, as the main contributions appear to lie primarily in the design and integration of a domain-specific modeling workflow. Given these additional results, I will raise my score from 1 to 2.

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

The paper proposes a requirement-aligned LLM framework for constructing HRSP optimization models from natural-language descriptions and instance data. It organizes expert formulations into reusable decision, constraint, and objective components, followed by component selection, component composition, and model and solution verification. A knowledge base built from 20 studies is evaluated on five held-out HRSP formulations and 15 single-requirement changes. Reported F1 scores are 0.952 for component selection and 0.966 for adaptation, with all returned schedules passing independent target-constraint checks.

Critical strengths
  1. The distinction between solver feasibility and target feasibility addresses a meaningful failure mode: executable formulations can omit production requirements. Independent target-constraint checks make the evaluation more informative than solver execution alone.

  2. Components with explicit production semantics and dependency records provide a structured approach to localized adaptation. The requirement-change experiments offer encouraging evidence that the framework can preserve unaffected components while making necessary updates.

Critical weaknesses
  1. The motivation does not sufficiently distinguish HRSP-specific challenges from general LLM-based optimization modeling challenges. Structural reformulation, requirement omissions, and formulation inconsistency also arise in general MILP modeling. These limitations also apply to the proposed framework, so the manuscript should more clearly establish its distinctive contribution and the scope of its reliability improvement.

  2. Adaptability depends strongly on knowledge-base coverage. Unresolved requirements prevent model acceptance, while constructing new components is deferred to future work. This limits applicability to customized HRSP production rules that cannot be expressed through existing components and supported derivations. The unclear human effort required to construct, verify, and extend the knowledge base further complicates the assessment of practical applicability.

  3. The experimental scope is narrow: only five held-out HRSP formulations are evaluated, and the 15 single-requirement changes derive from these same cases. This provides limited evidence of robustness across heterogeneous production settings. The number and scale of numerical instances associated with each formulation are also insufficiently specified.

  4. The comparisons do not isolate the contribution of Model and Solution Verification and its feedback-based correction from those of component selection and composition. Simpler baselines with comparable verification and correction are not evaluated, and corresponding ablations are absent. Shared generation settings and Pareto-search budgets do not establish comparable modeling-time computation; LLM calls, correction iterations, and modeling costs are not reported.

Questions for authors
  1. Which HRSP-specific modeling challenges motivate the proposed design beyond general LLM-based MILP modeling, and which aspects of the framework specifically address them?

  2. What can the current implementation do when a target requirement cannot be expressed through existing components and supported derivations? Which knowledge-base construction and verification steps require human experts, and how much effort is needed to accommodate a new production rule?

  3. Which studies constitute the five held-out formulations, and how many numerical instances of what sizes are evaluated for each? Are results available across repeated generation runs or alternative held-out selections?

  4. Do the baselines receive comparable verification feedback and opportunities for correction during model construction? What controlled results, if any, separate verification-related gains from component selection and composition under comparable modeling-time budgets, and what are the associated LLM calls, correction iterations, and modeling times?

2 comments

Thank you for recognizing the value of independent evaluation checks and adaptation to production requirements.

[W1] [Q1] HRSP modeling challenges

The central challenge in HRSP is not merely generating a mathematically executable optimization model but ensuring that multiple interdependent production decisions collectively represent the intended manufacturing process.

Unlike simpler modeling tasks in which constraints can often be interpreted relatively independently, HRSP involves tightly coupled slab selection, rolling-round assignment, sequencing, and production restrictions. A modeling error in one relationship may affect several other decisions, even when the resulting formulation remains mathematically feasible.

Our framework is designed to address this requirement-consistency challenge. Its contribution is to explicitly check whether production requirements are sufficiently represented and whether the mathematical relationships among the corresponding decisions are enforced. This provides a systematic way to identify and correct modeling errors that may not be revealed by solver feasibility alone.

MethodInitial F1Change precisionChange F1
DeepSeek+RAG0.9080.2500.394
ORThought0.8680.1490.255
MIRROR adaptation0.8340.1740.286
OptiMUS0.8620.2500.373
CoE0.8760.1710.286
AlphaOPT0.8790.2000.320
Ours0.9521.0000.966

F1 measures component-level modeling accuracy by balancing precision and recall. Our method achieves a change F1 of 0.966 with precision of 1.000, indicating accurate component updates without unsupported changes. This precision matters in HRSP: a local adjustment must be reflected in the affected decisions while preserving unaffected production rules.

[W2] [Q2] Automated knowledge-base construction

The agent extracts formulas with their explanations and source locations, groups them into decision, constraint, and objective components, and records production meanings and formula-supported connections. Programmatic validation checks formula ownership, source references, and dependency consistency. Expert knowledge is supplied by the source formulations; component grouping and connection recording are performed automatically.

[W2] [Q2] Unsupported requirements

When a target requirement cannot be expressed through existing components or supported derivations, the current implementation records it as unresolved and prevents model acceptance. Extending the library requires an additional source formulation or an expert-defined production rule, together with corresponding evaluation criteria. Experts clarify ambiguous rules or formalize missing production knowledge; the agent then extracts and organizes the resulting components and connections.

[W3] [Q3] Test instances and reproducibility

Thank you for requesting these details. We evaluate one numerical instance from each held-out study:

StudySlabsRolling rounds
Jiang et al., 2022, Comput. Ind. Eng.1506
Jiang et al., 2025, IEEE TASE1474
Liu et al., 2019, J. Ambient Intell. Humaniz. Comput.405
Zhang et al., 2020, IEEE TII1008
Jiang et al., 2025, Swarm Evol. Comput.1474

The reported results use one fixed held-out split. We will add repeated model-generation runs and evaluations using alternative held-out selections to further assess robustness.

[W4] [Q4] Validation and evaluation

ORLM, LLMOPT, ORThought, MIRROR, OptiMUS, CoE, and AlphaOPT retain the validation and correction procedures of their respective implementations. DeepSeek and DeepSeek + RAG do not receive our component-level validation feedback. In our framework, uncovered requirements trigger selection review, while dependency or connection errors trigger composition correction. All methods undergo the same evaluation checks after solving; ground-truth formulations and check outcomes are not fed back to generation or search.

[W4] [Q4] Ablation study

We report two additional removal variants alongside the previously reported Full and DeepSeek + RAG results:

ConfigurationInitial PInitial RInitial F1Change PChange RChange F1Stability
Full0.9940.9200.9521.000.9330.9661.00
DeepSeek + RAG0.9140.9070.9080.2500.9330.3940.917
w/o coverage recovery0.7400.4180.5320.1780.5330.2670.924
w/o decision-dependency feedback0.7360.4080.5220.1950.5330.2860.941

Removing coverage recovery reduces change F1 from 0.966 to 0.267, while removing decision-dependency feedback reduces it to 0.286. Coverage recovery identifies requirements and components omitted during initial selection, whereas dependency feedback identifies decisions needed to establish consistent relationships among the selected components. The performance deterioration when either mechanism is removed supports the importance of explicitly checking requirement completeness and mathematical dependencies.

I thank the authors for their response. In my view, the current version of the paper still has non-negligible shortcomings in its motivation and methodological design, so I will keep my original score.

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

The paper presents a requirement-aligned LLM modeling framework for hot-rolling production scheduling. It organizes modeling knowledge from prior studies into components with semantic descriptions and dependencies, combining requirement matching, coverage checking, component composition, and solution validation to support model generation and local requirement updates. Evaluation on five held-out cases and 1515 single-requirement edits shows fewer unrelated changes and higher target-constraint satisfaction rates in the tested cases, indicating the potential of structured knowledge reuse for domain-specific modeling.

Critical strengths

The paper addresses a practically relevant challenge in industrial optimization-model maintenance: ensuring that models faithfully capture production requirements and that requirement changes do not introduce unrelated modifications. Its representation of modeling knowledge as components with explicit semantics and dependencies provides an interpretable, inspectable foundation for knowledge reuse, requirement checking, and local edits. The evaluation is also carefully designed and interpreted: test studies are entirely excluded from the knowledge base, solver feasibility and target feasibility are reported separately, and tradeoffs across Pareto metrics are transparently discussed without presenting local gains as evidence of universal superiority over the reference method or feasible-region equivalence.

Critical weaknesses

The paper's methodological advance and industrial applicability remain insufficiently established. The framework combines a structured knowledge base, retrieval, LLM judgments, and repair, but does not clearly identify which requirement-alignment capabilities extend beyond existing knowledge-enhanced modeling methods such as ORThought, MIRROR, and OptGraph. Building the component library and independently validating requirements also depend on expert knowledge, while the process for supplying components and validation criteria for new rules remains unclear, limiting the evidence for sustained reductions in modeling and maintenance effort beyond the studied domain. The experimental design further obscures the sources of improvement: relative to full-text RAG, the method simultaneously changes knowledge granularity, manual curation, dependency graphs, coverage checking, and repair. Controls using the same component library, module ablations, and comparisons with closer methods are therefore needed to separate the contributions of knowledge resources and algorithm design; the role-based breakdowns in Tables 4 and 5 do not provide this separation. Finally, although the target checker appears in both the method and evaluation descriptions, its use for feedback and filtering, as well as baseline access, remains unclear. Consequently, the 100%100\% result in Table 2 does not clearly distinguish initial modeling ability from repair gains and filtering effects. These ambiguities raise questions about comparison fairness but do not establish leakage.

Questions for authors
  • Under the same component library, base model, call budget, and validation access, how does the method compare with ordinary RAG and MIRROR/ORThought? What are the individual contributions of the structured library and each module?
  • Please specify the target checker's source, invocation timing, feedback access, and the number and pass rate of plans before and after validation. For new requirements without an expert reference model, who constructs this checker?
  • Table 1 reports initial recall of 0.9200.920, while Table 5 reports stability of 1.0001.000 for every role. Under Equation (17), any initially omitted component that should remain after an edit lowers stability. Could you provide the sets for each case and clarify whether both experiments reuse the same initial output or involve additional repair or regeneration?
2 comments

Thank you for recognizing the practical relevance of our work.

[W] Primary contribution

We would first like to clarify that our framework and conventional RAG address fundamentally different tasks. RAG retrieves potentially relevant modeling knowledge, but relevance alone does not establish whether the retrieved formulations are applicable, whether all necessary modeling components have been included, or whether their mathematical relationships are correctly enforced.

Our contribution is to make this requirement-to-formulation correspondence explicit and verifiable. Rather than directly assembling retrieved formulations, our framework systematically examines what should be included, what may have been overlooked, and whether the assembled components jointly represent the intended production decisions.

This distinction is particularly important when requirements change, as the framework must identify necessary mathematical modifications while avoiding changes unsupported by the new requirements.

[Q1] Expanded comparisons

We appreciate the recommendation to include more closely related modeling systems. We have added ORThought, MIRROR, OptiMUS, CoE, and AlphaOPT. Our method achieves the highest initial-selection and adaptation F1 among all compared methods:

MethodInitial F1Change precisionChange F1
DeepSeek + RAG0.9080.2500.394
ORThought0.8680.1490.255
MIRROR adaptation0.8340.1740.286
OptiMUS0.8620.2500.373
CoE0.8760.1710.286
AlphaOPT0.8790.2000.320
Ours0.9521.0000.966

These baselines do not reliably identify the scope of a requirement change: their change precision is at most 0.250. Our precision of 1.000, with change recall of 0.933, shows that our method identifies necessary updates without unsupported changes.

[Q1] Bidirectional component selection

Ordinary RAG retrieves relevant formulations; our method checks whether they implement the target production requirements. Consider an HRSP model involving slab selection, rolling-round assignment, and sequencing. Retrieving valid components for these decisions individually does not ensure that they describe the same production plan. Without appropriate connections, an assigned slab may be omitted from the rolling sequence even though the MILP remains solver-feasible.

Forward matching checks each component’s production meaning and applicability against the target requirements, recording evidence for acceptance or rejection. Reverse review checks whether every requirement has a corresponding component, revisits rejected candidates, and rereads the original description for omissions. For example, if sequencing is required but no sequencing component was selected, this gap triggers reconsideration.

Composition then checks dependencies among the selected components. Missing decision dependencies return to selection; for declared decision links, solver-generated counterexamples expose inconsistent assignments and guide completion of missing connections.

[Q1] Contributions of the library and modules

The methodological contribution of our framework is a requirement-grounded modeling process that addresses three distinct sources of formulation errors: unsupported component inclusion, incomplete requirement representation, and inconsistent mathematical composition.

The structured library provides the semantic and mathematical information needed to assess component applicability. Bidirectional selection addresses the first two errors by examining both the justification for selected components and the coverage of target requirements. Executable composition addresses the third by testing whether declared relationships among production decisions are enforced. Importantly, these mechanisms work together to establish and maintain the correspondence between operational requirements and mathematical formulations, which differs from treating model construction as a direct consequence of knowledge retrieval.

The experimental results further support this distinction. DeepSeek + RAG achieves the same change recall as our method (0.933), indicating that retrieving relevant knowledge can recover many necessary modifications. However, its change precision is only 0.250, compared with 1.000 for our framework. This substantial difference suggests that the main advantage is not merely finding relevant modeling knowledge, but accurately determining which mathematical components should be modified while avoiding unsupported changes. This capability is particularly important in industrial scheduling, where unnecessary constraints can alter the intended optimization problem.

[Q2] Modeling checks and ground-truth evaluation

We agree that their sources, timing, and feedback access need to be stated explicitly, and appreciate the reviewer highlighting this ambiguity. Modeling validation guides correction during component selection and composition. It uses the target description, component definitions, and recorded dependencies. Requirement interpretation involves LLM judgments, while dependency tracing and solver searches for counterexamples are programmatic. Uncovered requirements trigger selection review; inconsistent dependencies or connections trigger composition correction. Baselines retain their own correction procedures and do not receive our component-level validation feedback.

Evaluation checks independently assess the final schedules against ground truth. We implement the held-out studies’ original formulations, fix each returned schedule’s production decisions, and check feasibility after solving. These checks apply to all methods and provide no feedback to generation or search.

Without a reference formulation, evaluation checks must be constructed from expert-defined production rules, consulting domain specialists where meanings are ambiguous. The requirement-change tests evaluate component updates, not solution feasibility using newly constructed checks.

[Q3] Stability of initially correct components

We have revised Eq. (17) to measure preservation among initially correct components unaffected by the requirement change. Let HiH_i denote the unchanged reference set and U^io,U^im\widehat{U}_i^o,\widehat{U}_i^m the paired predictions. The eligible set is Bi=Hi∩U^ioB_i=H_i\cap\widehat{U}_i^o, and Si=∣Bi∩U^im∣∣Bi∣S_i=\frac{\left|B_i\cap\widehat{U}_i^m\right|}{\left|B_i\right|} for nonempty BiB_i. Stability of 1.000 means that all initially correct, unaffected components were retained; initial recall of 0.920 separately measures initial coverage. We have also updated the corresponding table explanations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.