acceptodds
Under review as a conference paper at ICLR 2027

FROM-ROUTE: From Operations Research Problems To MIP Formulations For Better Solver Performance

Abstract

Formulating real-world operations research problems as mixed-integer programs (MIPs) requires choices of variables, constraints, and representations that shape solver performance. We present FROM-ROUTE, a framework for enhancing LLM routing among MIP formulations to improve solution performance within a fixed program budget. Deterministic analysis connects the original problem's structure and numerical data to three natural-language mechanism cards, supporting an ordered selection from an existing formulation-expert library. Multicommodity network design is the primary study of routing mechanisms; vertex coloring and resource-constrained project scheduling explore cross-domain applicability through domain-specific adapters. On 2,335 instances evaluated through independently validated trajectories, mean normalized primal integral is 57.0%, 14.6%, and 3.6% lower than that of the best fixed expert in each respective domain, alongside more validated model completions. Knowledge comparisons favor instance-conditioned cards over absent or generic mechanism context; restricted-completion relay is explored only for network design. On network design and coloring, FROM-ROUTE also improves solution metrics over direct and expert-guided modeling-kernel generation systems, using fewer provider-reported tokens in the calls producing retained decisions or code. These are descriptive comparisons: the evaluation corpora informed method development, and generation systems use different model and inference configurations.

Primary area:
optimization
Keywords:
Mixed-Integer Programming, Formulation Selection, Large Language Models, Retrieval-Augmented Generation, Algorithm Selection, Operations Research

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_E7vNOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes FROM-ROUTE, an LLM-based framework for selecting mixed-integer programming (MIP) formulations from a predefined library of formulation experts. The framework first extracts instance-specific structural and numerical features, then uses deterministic heuristics to construct mechanism-oriented knowledge cards. An LLM subsequently selects an ordered portfolio of three candidate formulations, which are executed under a shared computational budget. The resulting solutions are independently validated against the original optimization problems.

The framework is evaluated on three classical operations research problems: multicommodity network design (MCND), vertex coloring (VCP), and resource-constrained project scheduling (RCPSP), covering 2,335 instances and 40 formulation configurations. The authors report reductions of 57.0%, 14.6%, and 3.6% in mean normalized primal integral relative to the best fixed formulation in the respective domains, along with improvements in validated model-completion counts. Additional experiments compare different knowledge configurations, investigate a restricted-completion relay strategy for MCND, and evaluate the framework against LLM-based modeling-kernel generation systems.

Critical strengths
  1. Interesting and Significant topic: The topic of this paper is to auto-selecting MIP formulation for solving efficiency by LLM, which have real impact and could be further explored.
  2. Comprehensive dataset and reasonable metrics: This paper has 3 different datasets and use multiple metrics, such as objective value, constraint violation and primal integral, which are all effecitve metrics in field of solver developing.
Critical weaknesses
  1. Paper writing is terrible: This paper is quite long and includes all details, which is good, but it fails to show the core idea and has a lot of defensive expressions. These make this paper more like a report instead of a mature paper. I think the reason is, as the authors said, this paper was largely finished by Gen AI. Although I agree with using Gen AI to draft papers, which allows authors to spend more time on research, I believe the paper should be modified by the authors to make the storyline clear, instead of putting everything in the paper and letting readers spend too much time.
  2. Lack of baselines: I think a strong baseline in this setting is traditional machine learning model and deep learning based method, as in this working, the feature of problem has already been extracted. However, this paper only compare their method with LLM-native method, which cannot convince me. If traditional method is enough, why we use expensive LLM.
  3. I think the role of deterministic heuristic is unclear, which is complex and designed manually. Also in different datasets, the heuristic are complete different, which let me doubt that these heuristic is overfitted on each dataset. Besides, I think the ablation of heuristics is also weak, the execution of FROM-ROUTE can execute multiple formulations but Rule Top-1 can only execute one formulation.
Questions for authors

Please see the weakness section.

2 comments
AuthorsOpenReview ↗

Response to Reviewer E7vN

W1

We have rewritten the Introduction's method overview around how instance data and modeling experience inform LLM formulation selection. Instance analysis supplies structural and construction information, mechanism cards store modeling experience, and the LLM uses both to order experts. The commodity-aggregation example explains why model size and relaxation strength must be considered together for an instance.

Figure 2 shows these relationships and the information or decision step lost when a component is removed. The new paragraphs replace the brief overview and consolidate adjacent repetition; detailed interfaces and execution protocols remain in the method section and appendices. The authors are responsible for the final argument, text, and experimental interpretation, with the tools' supporting roles recorded in the AI Use Statement.

W2

We compare against all 16, 8, and 16 fixed expert configurations for MCND, VCP, and RCPSP, respectively, plus the non-LLM MCND Rule Top-1 replay. Figure 4, Table 3, and Appendix E.6 report these results. Fixed experts test the value of instance-level selection; rule replay tests direct top-recommendation execution; RAW compares systems that generate modeling kernels.

The LLM interprets natural-language modeling experience alongside numerical features, without training a task-specific selection model. For example, it uses commodity structure and construction size to interpret a card's explanation of aggregation's size/relaxation trade-off. Table 4 tests whether this knowledge improves decisions through the same LLM interface.

MCND Full uses an average of 2,540.94 tokens per instance in retained decision calls (Appendix E.4). We have not compared its performance and cost with a trained selector, so these tokens do not establish a cost advantage over traditional selection.

W3

The Analyzer extracts instance facts, and the recommender uses them to select relevant mechanism cards; the LLM determines the final expert order. MCND presents cards in name order, hides recommendation scores and ranks, and leaves the entire eligible library selectable. First choices fall outside the card-covered set on 20 of 196 instances (Appendix D.10). VCP/RCPSP also provide ranking priors that the LLM can revise.

Domain-specific analyses reflect different modeling objects, such as commodity aggregation, color representations, and time or resource representations. Recommendation parameters were adjusted using the reported benchmark results (Appendix D.3); fixed-parameter performance on independent instances has not been evaluated.

Rule Top-1 runs one expert, while Full may continue with others, so their difference compares complete systems. MCND-clean (Appendix F.6) compares knowledge contexts with the same shortlist, LLM configuration, and single-expert execution, without fallback or relay. Against No RAG and Generic RAG, Full reduces NPI by 11.3% and 12.7%, with original-problem completions increasing from 141 and 140 to 143.

Thank you for clarifying. I decide to keep my score.

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

The paper proposes an automatic LLM-based formulation of some operations research problems based on fixed templates + numerical data of the instance.

Critical strengths

None

Critical weaknesses

The writing is extremely obfuscated, with the introduction of new jargon at multiple places in the text. A notion of budget is mentioned early in the manuscript, and one has to look up section 4.4. before understanding that this is a control of the runtime of the MIP model. There is a similar case for what is implied by semantics in this article. NPI is not defined in text, one has to infer that this is a normalized version of the primal integral of Berthold 2013.

More fundamentally, it is not clear what problem the paper is solving. The paper does not provide insight into how to provide better formulations, or why/whether other baselines (what is RAW) fare worse.

Questions for authors

What is the question the paper wants to address, and is the goal to propose a method, or to demonstrate the capabilities of current LLM systems?

1 comment
AuthorsOpenReview ↗

Response to Reviewer zCuS

Weaknesses, paragraphs 1 and 2 / Q1

We have rewritten the Introduction (pages 1 and 2, marked in red) to explain the research question and the method's contribution. We study how numerical instance evidence and modeling experience can help an LLM select a MIP formulation that improves solver performance within a fixed program budget. The contribution is a formulation-selection method, evaluated through the solver outcomes of its choices.

The first added paragraph explains the components' responsibilities. Instance analysis supplies structural features, resource relationships, and construction-size estimates. Mechanism cards organize modeling experience as reusable natural-language knowledge about applicability and trade-offs. The LLM interprets this knowledge using the current instance's facts and produces an expert execution order. For example, commodity aggregation reduces model size but can weaken the relaxation; the choice depends on the instance's commodity structure and construction demands.

The second added paragraph and Figure 2 (page 3) explain how the components work together. Without mechanism cards, the LLM lacks the explicitly organized modeling experience; without instance analysis, that experience lacks structured numerical grounding in the current input. Replacing the LLM requires a separate rule or learning procedure to turn those inputs into a choice. These passages explain the design before introducing its formal definitions and implementation.

FROM-ROUTE makes this choice within an existing executable formulation library before solving. Section 3.1 and Appendix A describe the candidate formulations; Figure 4 and Table 3 compare FROM-ROUTE with fixed experts. The knowledge-context comparisons and MCND-clean test whether organizing the information supplied to the LLM improves its choices.

Weaknesses, paragraph 1

  1. Program budget covers expert startup, preprocessing, model construction, solving, solution recovery, and output. It is 3,600 seconds for MCND/VCP and 120 seconds for RCPSP. Analyzer computation, LLM requests, and external validation are accounted for separately (Section 4.4 and Table 2).
  2. Semantics comprises the original decisions, constraints, and objective, plus the mapping from model solutions back to those decisions. For example, MCND path variables are recovered as commodity flows on the original network and checked against its constraints. Section 3 formalizes this relationship through the feasible set, objective, and recovery map.
  3. NPI measures solution quality over the full budget and is defined in Section 5.2 as

NPIi,m=1T∫0Tℓi,m(t),dt.\mathrm{NPI}_{i,m}=\frac{1}{T}\int_0^T\ell_{i,m}(t)\\,dt.

Here, ℓi,m(t)\ell_{i,m}(t) is the normalized loss of the best valid solution so far relative to the common best-observed reference. It equals 1 before the method produces a valid solution. Earlier solutions and smaller objective gaps reduce NPI; Appendix B.1 specifies the normalization and integration rules.

Weaknesses, paragraph 2

RAW Direct and RAW Expert use LLMs to generate modeling kernels within an execution harness; RAW Expert also receives modeling guidance (Section 5.3; Appendix H.1). FROM-ROUTE selects executable experts from an existing library of formulations.

Table 5 reports NPI values of 0.03888, 0.02825, and 0.006719 for RAW Direct, RAW Expert, and Full on MCND, respectively. The corresponding VCP values are 0.2786, 0.3911, and 0.04485; Full also has more model completions in both domains. These are system comparisons with different model and generation configurations, so they do not isolate a particular prompt, model, or component.

Reviewer_VYtLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents FROM-ROUTE, a framework for selecting among existing executable MIP formulations using instance evidence and expert mechanism knowledge. A deterministic analyzer extracts structural features, a domain-specific rule supplies three mechanism cards, and an LLM selects an ordered triple of formulation experts for budgeted execution. The framework is evaluated on multicommodity network design, vertex coloring, and resource-constrained project scheduling, using original-space solution validation, normalized primal integral, and validated model completions.

Critical strengths
  1. The paper addresses a practically relevant problem: formulation choices can substantially affect MIP performance under limited computational budgets.
  2. The framework organizes established modeling knowledge across three OR domains and provides a concrete interface for combining instance evidence, expert knowledge, and LLM-based selection.
  3. Independent validation of recovered original-space solutions is a useful design choice. The paper also explicitly distinguishes restricted-model completion from original-problem optimality and reports several limitations of its comparisons.
Critical weaknesses
  1. Generalization is insufficiently established. Appendix D.3 states that the card-recommendation coefficients were manually adjusted using results on the reported benchmark instances. The reported evaluation therefore does not clearly separate method development from testing. Furthermore, each domain requires its own features, scoring rules, and mechanism knowledge. The three-domain results demonstrate applicability with domain-specific engineering, but provide limited evidence of generalization to unseen instances or domains.
  2. The incremental contribution of the LLM remains unclear. The MCND Rule Top-1 replay is useful, but it does not match Full’s fallback and restricted-completion relay policy. The No-relay ablation also shows that execution policy materially affects the reported gains. A deterministic selector evaluated with the same execution policy is needed to isolate the value of LLM-based selection, particularly in VCP and RCPSP, where the LLM receives explicit ranking priors. The paper additionally lacks a controlled comparison across LLM backbones.
  3. The card design and its contribution are not sufficiently analyzed. Card construction depends on manually specified mechanisms and heuristics, but the choice of three cards is not supported by a card-count ablation. Existing context comparisons jointly change content, length, membership, and—in VCP/RCPSP—ranking priors. Although Appendix D.10 provides two brief decision examples, it does not establish how the cited mechanisms explain realized solver performance or when overriding the recommendation helps or hurts.
  4. The methodological novelty is limited in the current evidence. The framework combines established formulations, manually designed recommendation rules, LLM selection, and budgeted execution. This integration is useful, but the experiments do not yet establish a sufficiently generalizable selection principle or clearly isolate what new capability is contributed by the LLM and mechanism cards.
Questions for authors
  1. Can the authors clarify which instances informed the recommendation coefficients and prompt design, and provide evaluation on instances excluded from all such development?
  2. How does a deterministic selector perform under exactly the same fallback, relay, and execution budget as Full? For VCP/RCPSP, what happens when the supplied recommendation order is executed directly?
  3. How sensitive are the results to the number of cards and the LLM backbone? Can the contribution of mechanism text be separated from numerical ranking priors?
  4. Can the authors expand a decision example into a mechanism-level analysis linking instance features, card recommendations, the LLM’s choice, and the resulting solver trajectory, including an unsuccessful or harmful override?
1 comment
AuthorsOpenReview ↗

Response to Reviewer VYtL

W1 / Q1

Our study examines how instance data and modeling experience improve LLM formulation selection. Domains share the decision interface but supply their own structural features and mechanism knowledge. Commodity aggregation in MCND, color representations in VCP, and time and resource representations in RCPSP require different analyses. The revised Introduction and Figure 2 explain this interface; Table 4 and Appendix F.6 evaluate it through knowledge-context comparisons.

We manually adjusted recommendation coefficients using results on the reported benchmarks and fixed them before production decision calls (Appendix D.3). The manuscript does not report evaluation on instances excluded from development or a complete instance list for each prompt revision. The results therefore assess selection on the studied collections with domain-specific adaptations.

W2 / Q2

The LLM uses instance facts and mechanism knowledge to compare experts and determine their execution order. MCND-clean (Appendix F.6) tests how context affects these choices under single-expert execution. Its three conditions share the same instance-derived three-expert shortlist and LLM configuration. Only the first-ranked expert runs, with neither fallback nor relay.

MCND-clean conditionMean NPI ↓Original-problem completions ↑
Full0.01373836143/196
No RAG0.01548315141/196
Generic RAG0.01574314140/196

Full reduces NPI by 11.3% and 12.7% relative to No RAG and Generic RAG, with two and three additional completions, respectively.

Full versus No relay examines the relay policy; Rule Top-1 executes only the top recommendation, whereas Full may continue with other experts. The comparison with rule replay includes these differences in execution. We have not reported results for rules using Full's execution policy or directly executing the VCP/RCPSP recommendation orders. The context comparisons assess how to improve LLM selection; its incremental gain over rules requires those matched comparisons.

W2, W3 / Q3

Mechanism cards express modeling experience as natural-language knowledge about formulation trade-offs and applicability. The Introduction uses commodity aggregation as an example: smaller models can have weaker relaxations, a trade-off interpreted using commodity structure and construction size. The recommender supplies relevant cards for the LLM to read alongside instance facts.

Table 4 compares context packages whose content, method coverage, and length change together. MCND hides recommendation scores and ranks and presents cards in name order; its context gains occur without an explicit ranking prior. VCP/RCPSP also supply ranking priors; the comparisons do not isolate text content, length, card membership, or prior effects.

Three cards are an implementation choice; card-count and LLM-backbone sensitivity have not been tested.

W3 / Q4

Appendix D.10 records two choices made with cards covering DA, FA-O, and PAe-D-K1. For r01.2, the LLM selects PATH-K10 first, citing a path-construction proxy of 280 and noting that feasibility is unknown. For r01.1, it selects PAe-D-K1 first, citing proxy work of approximately 618.4 and retained distinction of approximately 0.738. These quantities are construction proxies, not measured runtimes or LP bounds.

Across 196 instances, 20 first choices fall outside the card-covered set. We selected the examples lexicographically, without consulting their solver outcomes. The requested paired trajectory analysis, including harmful overrides, has not been reported; the decision records alone do not establish when changing recommendations helps.

W4

FROM-ROUTE provides a method for selecting formulations before solving, using instance evidence and natural-language modeling experience to inform an LLM. The Analyzer extracts structural facts, mechanism cards explain relevant trade-offs, and the LLM uses both to order existing formulation experts. The revised Introduction and Figure 2 describe these responsibilities and their relationships.

Figure 4 and Table 3 test the resulting solver performance against fixed experts. Table 4 and MCND-clean test whether instance-conditioned mechanism context improves the LLM's choices.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.