acceptodds
Under review as a conference paper at ICLR 2027

Post-Training Open Multimodal Models for Optimization Modeling from Text and Images

Abstract

Large language models (LLMs) have advanced automated optimization modeling, but existing methods largely assume that problems are fully specified in text. In practice, text and images often provide complementary information needed to formulate a problem. While existing multimodal optimization modeling benchmarks reveal a substantial performance gap between open and proprietary multimodal large language models (MLLMs), they do not provide training data or post-training methods for this task. We therefore ask whether task-specific post-training can equip a compact open MLLM with this capability. We introduce MLLM4OR-Corpus, comprising 22,107 solver-verified instances. We partition the corpus into disjoint training, development, and test subsets for post-training, model selection, and in-domain evaluation, respectively. For external evaluation, we use MM-OptBench and introduce LLM4OR-VL to cover a wider range of problem sources and description styles. LLM4OR-VL comprises 878 multimodal problems reconstructed from eight text-only optimization benchmarks. Using these resources, we introduce MLLM4OR, an open 9B-scale multimodal LLM trained to generate mathematical formulations and corresponding solver code from multimodal specifications. To this end, we first use supervised fine-tuning (SFT) to connect text-image evidence to formulations and solver code, comparing training answers with no explanations, concise explanations, or detailed explanations of the modeling and implementation steps. We then propose a solver-guided preference construction method for direct preference optimization (DPO), contrasting verified answers with the model's own execution and modeling failures. Extensive experiments show that including concise explanations for some training problems and omitting them for others yields the strongest overall SFT performance. DPO further improves overall accuracy and performance on MM-OptBench and LLM4OR-VL. These findings support task-specific post-training for automated multimodal optimization modeling with compact open MLLMs.

Primary area:
optimization
Keywords:
Optimization Modeling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_7QsUOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents MLLM4OR, a 9B multimodal large language model designed to translate complementary text-image optimization specifications into mathematical formulations and executable solver code. The authors construct MLLM4OR-Corpus, a 22,107-problem multimodal optimization corpus, and investigate several supervised fine-tuning strategies, including Direct-SFT, Compact-SFT, Detailed-SFT, and a Mixture-SFT strategy.

The paper further proposes SG-DPO, a solver-grounded direct preference optimization procedure. The central idea is to construct preference pairs using model-generated candidate solutions that are audited through execution and solver-based checks, with verified reference solutions serving as positive responses. The final preference dataset contains 498 preference pairs from 315 training problems.

The resulting model achieves strong performance on the proposed MLLM4OR-Test benchmark and demonstrates meaningful cross-dataset generalization on LLM4OR-VL and MM-OptBench. In particular, Mixture-SFT reaches 91.0% on MLLM4OR-Test, while SG-DPO obtains 49.5% on LLM4OR-VL and 44.3% on MM-OptBench, outperforming the corresponding Mixture-SFT models on the two external benchmarks. The paper also reports Pass@4 results and comparisons with several strong general-purpose multimodal models.

Overall, the paper addresses an interesting and practically relevant problem at the intersection of multimodal reasoning, mathematical optimization, structured generation, and solver-based verification.

Critical strengths
  1. Substantial task-specific multimodal corpus and end-to-end task formulation.

A major strength is the construction of a relatively large 22,107-problem corpus for a difficult and underexplored task: translating complementary text-image optimization specifications into mathematical formulations and executable solver code. The work goes beyond generic visual question answering by requiring structured mathematical reasoning and executable outputs.

The combination of multimodal inputs, mathematical formulation, and solver execution creates a practically meaningful benchmark for studying reliable multimodal structured generation.

  1. Strong empirical performance and evidence of cross-dataset generalization.

The paper provides a broad empirical evaluation rather than relying only on the proposed benchmark. Mixture-SFT substantially improves performance over the baseline on MLLM4OR-Test, and SG-DPO provides additional gains on LLM4OR-VL and MM-OptBench. The external benchmark results are especially valuable because they suggest that the learned capabilities are not completely confined to the training distribution.

The use of solver-based verification and Pass@4 analysis also provides a more informative evaluation protocol than simple language-generation metrics.

Critical weaknesses
  1. The methodological novelty of SG-DPO is not yet sufficiently isolated.

The core DPO objective is standard, and the main methodological contribution is the solver-grounded construction of preference pairs. However, the current experiments do not establish that solver-grounded preference construction itself is responsible for the observed gains.

For example, it would be important to compare SG-DPO against equally sized preference datasets constructed using: (a) random model-generated negatives, (b) hard negatives without solver verification, and (c) rejection-sampled supervised fine-tuning or another supervised approach using the same verified/generated candidates. Without such controls, the improvement could plausibly result from additional targeted training or hard-negative exposure rather than the specific DPO formulation.

This is the main issue affecting my assessment of the paper's methodological novelty.

  1. The multimodal contribution and semantic fidelity are not sufficiently isolated.

Although the task is multimodal, the paper does not appear to provide direct controlled evidence quantifying how much the image contributes to the final solution. Text-only, image-only, and image-shuffled/corrupted-image experiments would substantially strengthen the claim that the model learns genuine multimodal grounding.

Moreover, successful solver execution does not by itself guarantee semantic fidelity. A formulation may be syntactically valid and solvable while encoding an incorrect parameter, variable domain, objective function, or individual constraint. A more granular evaluation of these semantic components would therefore provide stronger evidence for the claimed text-image-to-optimization capability.

The source composition of the evaluation set is another limitation: approximately 85.6% of MLLM4OR-Test is derived from OptMATH, so the headline aggregate score is heavily influenced by this source. In addition, the external LLM4OR-VL benchmark is reconstructed from text-only optimization sources, which makes it useful for cross-source generalization but not a direct test of naturally occurring multimodal optimization problems.

Questions for authors
  1. Can you isolate the specific contribution of solver-grounded preference construction in SG-DPO?

In particular, could you report controlled comparisons using the same number of preference pairs and the same initialization, with negatives constructed by:

random model-generated responses; hard model-generated responses without solver-based verification; solver-verified hard negatives but trained with a supervised/rejection-sampling objective instead of DPO?

This would clarify whether the gains come specifically from solver-grounded preference optimization or primarily from additional targeted training on difficult examples.

  1. How much of the final performance actually depends on the visual input?

Could you provide text-only, image-only, and image-shuffled/corrupted-image ablations under the same training and evaluation protocol? This would help establish whether the model is genuinely using visual information to resolve optimization specifications, rather than relying predominantly on the textual description.

  1. What exactly is checked by the solver-based correctness evaluation?

Could you clarify whether correctness is evaluated at the level of individual numerical values, variable names and domains, objective expressions, and individual constraints, or whether the primary criterion is successful execution/solver completion? If the latter, can you provide a semantic-fidelity analysis or component-level error breakdown?

  1. How robust are the reported gains to the source imbalance in MLLM4OR-Test?

Since approximately 85.6% of the test set is derived from OptMATH, could you report source-balanced or per-source aggregate metrics, such as macro-average accuracy across OptMATH-derived and OR-Instruct-derived subsets? This would help determine whether the overall 91.0% score reflects broad generalization or is primarily driven by the dominant source.

Reviewer_j8HTOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces LLM4OR, a 9B multimodal model post-trained to generate mathematical formulations and executable optimization code from combined text and image inputs. The authors construct a 22K-instance corpus and a separate 878-instance benchmark, LLM4OR-VL. They investigate different levels of reasoning supervision during SFT and introduce solver-grounded DPO using model-generated failures as negative examples. Experiments on three benchmarks show substantial improvements over the base model, with mixed reasoning supervision providing the best overall SFT performance and DPO further improving external benchmark results.

Critical strengths

This paper has three strengths:

  • A meaningful and relatively underexplored problem setting. The paper addresses a practical gap in existing optimization modeling research, which has largely focused on text-only specifications. Generating mathematical formulations and solver code from complementary visual and textual information is a useful task, and the paper makes a reasonable case for its importance.

  • Substantial effort in dataset construction. MLLM4OR-Corpus contains 22107 instances with reference formulations and solver code, supported by human review and solver-based verification. The additional LLM4OR-VL benchmark also broadens the evaluation sources. These resources are potentially valuable contributions on their own.

  • Strong improvements from task-specific post-training. The improvements over the untuned Qwen3.5-9B are substantial. For example, performance on MM-OptBench increases from 11.8% to 44.3%, while LLM4OR-VL improves from 28.0% to 49.5%. The gains on external benchmarks make the results more convincing than improvements on the authors' own test set alone.

  • A practical use of solver feedback for preference learning. The distinction between execution failures and programs that execute successfully but produce incorrect objective values is sensible. Constructing preference pairs from the model's actual errors is a practical approach, and the additional failure analysis provides useful insight into the remaining limitations.

Critical weaknesses

I have the following concerns:

  • The contribution of visual understanding is not sufficiently established. This is my main concern. Although the paper focuses on multimodal optimization modeling, the experiments do not directly quantify how much the trained model benefits from visual information or how well it actually interprets it. In particular, Appendix D.6 indicates that complete text-only versions of LLM4OR-VL already exist, but they are not used for comparison. A text-only oracle baseline, together with controlled image perturbations, would help separate visual understanding from downstream optimization reasoning. This is particularly relevant given the concerns raised in [1].

  • The evaluation does not fully establish the correctness of generated formulations. The main metric relies primarily on matching the reference optimal objective value. However, two different formulations can produce the same optimum for a particular instance while representing different feasible regions or constraints. The generated mathematical formulation is also not directly checked for consistency with the executable code. The authors acknowledge this limitation, but it remains important because producing faithful formulations is a central claim of the paper. Prior work such as [2,3] highlights this issue.

  • The extent of generalization beyond reconstructed and template-based problems remains unclear. Most examples are reconstructed from existing text-only problems or generated from structured templates, with images rendered programmatically. Moreover, the parametric generators are inspired by the problem families in [4], and the paper acknowledges that similar templates or source material may occur across datasets. The large gap between in-domain and external performance further motivates a stronger examination of generalization. The current results demonstrate adaptation to the evaluated tasks, but provide less evidence about robustness to genuinely new problem structures or naturally occurring visual specifications.

  • The explanation-supervision experiments do not fully isolate the effect of reasoning supervision. The comparison between Direct-SFT, Compact-SFT, Detailed-SFT, and Mixture-SFT is interesting, but the variants differ in supervised-token counts and training instructions, and only one training seed is reported. The paper also provides limited results showing sensitivity to the different mixture ratios. As a result, it is difficult to determine whether the improvement comes specifically from combining concise explanations with direct answers, rather than differences in training distribution or optimization dynamics. The conclusions about the optimal amount of explanation supervision seem somewhat stronger than the controlled evidence supports.

Refs:

  1. MathVerse: Does Your Multi-modal LLM Truly See the Diagrams in Visual Math Problems?
  2. Certified Evaluation for LLMs in Optimization Modeling: From Graph Isomorphism to Formulation Isomorphism.
  3. OptArgus: A Multi-Agent System to Detect Hallucinations in LLM-based Optimization Modeling.
  4. MM-OptBench: A Solver-Grounded Benchmark for Multimodal Optimization Modeling
Questions for authors

N/A

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

This paper studies task-specific post-training of an open multimodal model for translating complementary text-and-image specifications into mathematical optimization formulations and executable Gurobi programs. It introduces MLLM4OR-Corpus, containing 22,107 instances with 19,189 used for training, and LLM4OR-VL, an evaluation suite of 878 problems reconstructed from eight text-only benchmarks. Using Qwen3.5-9B, the authors compare direct, concise-explanation, detailed-explanation, and mixed SFT targets, followed by solver-grounded DPO that contrasts reviewed reference answers with policy-generated failures. Evaluation covers the held-out corpus split, LLM4OR-VL, and MM-OptBench through execution-based scoring and four-sample coverage. The reported results show substantial improvements over the untuned baseline, with SG-DPO reaching 90.1%, 49.5%, and 44.3% mean pass@1 on the three suites, respectively. Mixed supervision achieves the strongest aggregate SFT performance, while DPO improves external mean accuracy and sample coverage with a small in-domain decrease.

Critical strengths
  1. The corpus pairs complementary text-and-image specifications with mathematical formulations and executable reference programs. The construction process includes human checks of source requirements, visual information, and formulation–code correspondence, in addition to solver execution (§3.2; Appendix C). The documented two-reviewer assessment and adjudication process is a meaningful strength: the reference data are not accepted solely because a program runs. Providing dedicated training examples for this task could support research beyond the particular model evaluated here.
  2. On the existing MM-OptBench benchmark, Table 1 reports an improvement from 11.8% for the untuned backbone to 40.8% with Direct-SFT and 44.3% with SG-DPO. The source-, family-, and difficulty-level breakdowns provide evidence beyond a single pooled score. Although the baseline deployment differs from that of the trained models, the reported magnitude and breadth of the improvements make task-specific post-training a practically interesting direction.
Critical weaknesses
  1. Appendix E.1 states that Compact-SFT and Detailed-SFT are trained with instructions requesting an explanation before the formulation and code, whereas evaluation requests only the formulation and code. Direct-SFT and Mixture-SFT therefore receive training examples matching the evaluation format, while the explanation-only variants face an output-instruction shift. The comparison identifies a useful recipe for direct-output deployment, but it does not isolate the effect of explanation amount or frequency from instruction compatibility. The acknowledged differences in supervised-token budgets further limit this interpretation. For SG-DPO, all preferred responses are canonical targets already available during SFT, yet there is no matched continued-SFT control on the same selected problems or a simpler DPO control isolating the proposed preference construction. Appendix F.6 explicitly acknowledges the absence of component ablations. Consequently, the evidence supports the complete training recipe more clearly than the benefits of its particular design choices. This matters because those choices constitute much of the paper’s methodological contribution beyond the dataset.
  2. The stated task requires a faithful mathematical formulation and a corresponding implementation (§3.1), but the primary score mainly checks successful execution and agreement with a reference objective value (§B.2). A program can omit a constraint or change the feasible set without changing the optimum of the evaluated instance. The authors acknowledge this limitation, but no reported audit estimates how often objective-matching generated answers actually preserve the problem requirements and agree with their accompanying formulations. The semantic review of reference data does not resolve this question about model-generated outputs. The multimodal component is also insufficiently characterized. Appendix D.6 provides complete oracle-text counterparts for all 878 LLM4OR-VL problems, but the results do not use them to distinguish visual interpretation from downstream formulation and code-generation improvements. These omissions do not negate the reported end-to-end gains, but they leave the central modeling capability only indirectly validated.
Questions for authors
  1. Does the ranking of Direct-, Compact-, Detailed-, and Mixture-SFT persist when each model is evaluated with its training-consistent output instruction, allowing explanations where appropriate while scoring the resulting formulation and code? A comparison across direct-output and explanation-first inference, together with actual supervised-token counts and selected training steps, would help separate explanation supervision from output-format compatibility.
  2. How does SG-DPO compare with continued SFT on the same 315 selected problems and canonical preferred answers under a comparable additional-training budget? How much of the improvement remains with ordinary unweighted DPO on the same retained pairs? These controls would clarify the value of learning from rejected responses and of the proposed weighting.

Comments on the paper

1 comment
Area_Chair_ik4fOpenReview ↗

Author- Reviewer Discussion

Dear Reviewers, the authors have submitted their response. Kindly review it and update your review where appropriate.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.