acceptodds
Under review as a conference paper at ICLR 2027

MIND: Automated Optimization Modeling via a Localizable Error-Driven Perspective

Abstract

Automated optimization modeling via Large Language Models (LLMs) has emerged as a promising approach to assist complex human decision-making. While post-training has become a pivotal technique to enhance LLMs' capabilities in this domain, its effectiveness is severely constrained by the scarcity and underutilization of high-quality training data. However, through a detailed profiling of error patterns across various problem-response pairs drawn from post-training, we identify two fundamental limitations of existing automated optimization modeling approaches: (L1) the sparsity of error-specific problems and (L2) the sparse rewards associated with difficult problems. We demonstrate that these limitations can result in suboptimal performance in domain-specific post-training for LLMs. To tackle the above two limitations, we propose a novel error-driven learning framework—namely, automated optimization modeling via a localizable error-driven perspective (MIND)—that customizes the whole model training framework from data synthesis to post-training. MIND is based on our key observation of the unique localizable patterns in error propagation of optimization modelings, that is, modeling errors may remain localized to specific semantic segments and do not propagate throughout the entire solution. Thus, in contrast to holistic reasoning tasks such as mathematical proofs, MIND leverages the construction of a focused, high-density training corpus and proposes Dynamic Supervised Fine-Tuning Policy Optimization (DFPO) to tackle difficult problems through localized refinement. Experiments on six benchmarks demonstrate that MIND consistently outperforms all the state-of-the-art automated optimization modeling approaches. Furthermore, we will open-source the training dataset, code, and model for the community.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LLM post-training, automated optimization modeling

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_LsfHOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes MIND, a framework for automated optimization modeling based on the observation that errors in an LLM-generated optimization model are often localized to a subset of variables, constraints, or objective terms. MIND combines error-driven reverse data synthesis, which creates new optimization problems around identified failure patterns, with Dynamic Supervised Fine-tuning Policy Optimization (DFPO), which invokes a teacher model to correct failed rollout groups and jointly optimizes RL and supervised losses.

Critical strengths
  1. The paper explores an interesting perspective on automated optimization modeling: modeling errors may be localized to specific components while the remaining formulation retains its validity, motivating targeted error correction rather than complete regeneration.
  2. DFPO dynamically combines reinforcement learning with teacher-guided supervised fine-tuning to alleviate sparse reward issues when all sampled responses are incorrect. The extensive experiments and ablation studies provide empirical evidence supporting the effectiveness of the proposed framework.
Critical weaknesses
  1. The assumption of localized modeling errors is not sufficiently characterized or empirically supported. Errors in optimization modeling may propagate across components; for example, an inappropriate variable definition may lead to incorrect constraints. Figure 2 quantifies the number of erroneous components but does not establish whether errors in different components are independent or whether error propagation occurs.
  2. There is insufficient alignment between DFPO and its underlying motivation. While DFPO addresses sparse rewards in RL training, it remains unclear how the locality of modeling errors contributes to its effectiveness, as the SFT loss is computed over the entire corrected response rather than only the erroneous spans.
  3. The justification for the auxiliary reward RfidR_{fid} is insufficiently rigorous. Objective-value deviation may not reliably reflect formulation correctness: a minor constraint error can render a model infeasible, while a substantially incorrect formulation may yield a similar objective value by chance. Moreover, the paper does not specify how RfidR_{fid} is computed for infeasible or unbounded models that generated by LLM.
Questions for authors
  1. Does the proposed data synthesis and augmentation process account for error propagation across modeling components? If not, how do the proposed data synthesis method and DFPO handle such cases?
  2. Although Appendix C.3 examines controlled modeling errors, it remains unclear whether the assumed positive correlation between modeling error ratio and objective-value deviation also holds for errors in LLM-generated formulations. As a possible experiment, the authors could extend the analysis in Figure 2 by retaining the existing analysis of error ratios and additionally examining the relationship between the error ratios of LLM-generated formulations and their objective-value deviations from the corresponding ground-truth formulations. Furthermore, how is RfidR_{fid} defined when a generated model is infeasible or unbounded?
  3. During DFPO training, how are student responses selected for teacher-guided correction when all responses in a rollout group are incorrect? Does the teacher correct all failed responses or only a selected response, and, if the latter, what selection strategy is used? The paper does not clearly specify this procedure.
5 comments

Dear Reviewer LsfH,

We thank you for the constructive comments.

Q1: Error Locality and Propagation. (weakness 1 and question 1)

A1: Thank you for this question. We clarify the distinction between error propagation during generation and error locality in completed formulations, and explain how MIND uses this observation.

(1) Qualitative and quantitative evidence. Figure 1 provides a qualitative example in which an incorrect variable type leaves the parameters, objective function, and constraints valid. Figures 2 and 11 provide quantitative evidence from 100 failed responses per model across Qwen2.5-7B-Instruct, DeepSeek-V3, GPT-5, and Gemini-2.5-Pro. The mean proportion of erroneous components ranges from 28% to 33%.

(2) Error propagation and final error locality. During generation, an incorrect variable definition may affect dependent constraints. Our statistics characterize the resulting errors in completed formulations, where such dependencies can coexist with substantial valid content. This observed locality of the final errors motivates both reverse data synthesis and DFPO.

(3) Error locality in MIND. Reverse data synthesis uses collected error patterns to generate new problems and their solutions (Section 4.1). When all student rollouts in a group are incorrect, the teacher corrects the student responses using a reference formulation, allowing dependent components to be revised together while retaining valid content. A corrected response is eligible for SFT only if solver execution succeeds and its objective value matches the ground truth (Section 4.2 and Appendix B.4). Figures 16 and 17 illustrate successful local corrections on ComplexLP and OptMATH-Bench. Through dynamic SFT, DFPO uses these corrected responses to provide supervised learning signals for difficult rollout groups in which the binary accuracy reward provides no distinction among responses.

Q2: Error Locality and Supervision over Complete Responses. (weakness 2)

A2: Thank you for raising this point.

(1) Supervision over complete responses. DFPO builds on the locality of errors in completed formulations. Although errors may propagate during generation, correcting the erroneous components while retaining valid content produces a complete response that can serve as a supervised target. Under autoregressive next token prediction (Equation (4)), supervision over the complete response teaches the model both to generate repaired spans from preceding tokens and to generate valid continuations conditioned on the corrected prefix. Locality therefore guides target construction, while SFT learns the complete corrected response.

(2) Correction with reference context. Qwen3.8-27B supports nonthinking (nothink) and thinking modes, with low, medium, and extra high (xhigh) reasoning effort in thinking mode. We select the 82 of 166 OptMATH-Bench questions answered correctly by GPT-6-astra and compare two tasks while holding the mode and reasoning effort fixed. For direct answering, the model independently answers each original question using the paper's evaluation prompt, without a reference answer. For each incorrect response from the initial direct answering stage, we apply the paper's teacher correction prompts using the same model, mode, and reasoning effort. The model first generates a PySCIPOpt reference solution from the question and GPT-6-astra's full response. It then receives the question, its incorrect initial response, and the generated reference to repair the erroneous parts of that initial response while retaining valid content, and outputs a complete corrected response.

Temperature is 1.0 and the maximum generation length is 73,728 tokens. Both accuracies use all 82 questions and the same PySCIPOpt execution and objective value matching criterion. We compute similarity between each wrong response and its corresponding corrected response, and average the scores over successful corrections.

Reasoning effort is not applicable in nonthinking mode. Gain is measured in percentage points (pp). Tokens count API input and output, including reasoning tokens. Additional correction tokens cover both reference generation and repair, excluding direct answering.

ModelReasoning modeReasoning effortDirect response accuracyAccuracy after correctionGain (pp)Mean direct tokensMean additional correction tokensWord level F1 scoreEmbedding cosine similarity
Qwen3.8-27BNonthinkingN/A2/82 (2.44%)29/82 (35.37%)+32.936,113.0118,300.800.62280.9291
Qwen3.8-27BThinkingLow15/82 (18.29%)65/82 (79.27%)+60.986,384.8920,041.390.97240.9934
Qwen3.8-27BThinkingMedium17/82 (20.73%)68/82 (82.93%)+62.206,740.3421,452.150.94230.9901
Qwen3.8-27BThinkingExtra high23/82 (28.05%)66/82 (80.49%)+52.4427,532.3250,539.720.94830.9854

These results suggest that correcting an existing response with reference context is more tractable than direct answering in this setting, with successful corrections retaining substantial student content. The model uses task information supplied in context to repair its response, consistent with the context learning perspective in CL-bench [1].

(3) Training with corrected responses. Starting from the same Qwen3.8-27B base model, we train with DAPO and DFPO on the same 10K instances for two epochs. Both students use nonthinking mode, and DFPO uses the same backbone in thinking mode with medium reasoning effort as its teacher, with training reference formulations supplied as hints. Both students are evaluated in nonthinking mode under identical Pass@1 settings. Evaluation includes the six original benchmarks, ComplexOR (18 instances), and MIPLIB-NL (220 instances after excluding three instances with empty data files). Pass@1 accuracy (%) is reported below:

ModelNL4OptIndustryOREasyLPComplexLPOptMATH-BenchOptiBenchComplexORMIPLIB-NLMacro AVG
Qwen3.8-27B (DAPO)86.156.092.754.749.454.455.629.159.7
Qwen3.8-27B (DFPO)89.855.092.862.155.454.950.036.462.0

DFPO improves on six of eight benchmarks, supporting its effectiveness with supervision over complete corrected responses.

Q3: Modeling Fidelity Reward and Invalid Formulations. (weakness 3 and question 2)

A3: Thank you for this question. We agree that objective value deviation is an imperfect proxy for formulation fidelity. We use the modeling fidelity reward as an auxiliary signal for error severity alongside the binary accuracy reward. DFPO excludes infeasible or unbounded student formulations from reward computation: they receive neither accuracy nor modeling fidelity rewards. The formulations for all problems in the training set have finite optimal objective values verified by the solver. We use R=0.2Rfid+0.8RaccR=0.2R_{\mathrm{fid}}+0.8R_{\mathrm{acc}}, assigning the fidelity term a lower weight. Figure 13 and Appendix C.6 evaluate its training benefit using DAPO on the same 10K instances for seven epochs. Compared with accuracy reward alone, adding the fidelity term improves OptMATH-Bench accuracy from 9.0% to 17.5% and ComplexLP from 39.4% to 44.3%. These results demonstrate the training benefit of using the modeling fidelity reward as an auxiliary signal for error severity.

To further examine the statistical relationship, we repeat the experiment in Appendix C.3 using 100 training formulations generated by LLMs with correct objective values, including 50 with fewer than 20 variables and 50 with more than 20. We perturb kk of NN constraints through constraint removal, right hand side (RHS) modification, coefficient perturbation, or inequality direction flipping, generating 100 variants per perturbation level. Candidates that are infeasible, unbounded, or time out are excluded because a valid optimal objective comparison is unavailable. Within each formulation, we calculate Pearson and Spearman correlations between the constraint perturbation ratio E=k/NE=k/N and mean absolute objective deviation across perturbation levels, then average the correlations across formulations:

Formulation cohort# FormulationsPearson rSpearman rho
Variables <20500.9020.944
Variables >20500.6690.765
All formulations1000.7850.854

In the within formulations analysis, we calculate Pearson and Spearman correlations separately for each formulation using absolute objective deviation, then average the respective coefficients across the 100 formulations. In the cross formulations analysis, we first account for differences in objective scale using normalized deviation dnorm=∣Obj−Obj∗∣max⁡(∣Obj∣,∣Obj∗∣)d_{\mathrm{norm}}=\frac{\lvert\mathrm{Obj}-\mathrm{Obj}^{\ast}\rvert}{\max(\lvert\mathrm{Obj}\rvert,\lvert\mathrm{Obj}^{\ast}\rvert)}. We then combine the pairs of perturbation ratio and mean normalized deviation at each perturbation level from all 100 formulations into a single dataset and calculate its overall Pearson and Spearman correlations. The two analyses are summarized below:

AnalysisObjective deviation# FormulationsPearson rSpearman rho
Within formulationsAbsolute1000.7850.854
Cross formulationsNormalized1000.5660.603

The positive correlations within and across formulations indicate that objective value deviation increases with the proportion of perturbed constraints on average, supporting its use as an auxiliary signal for modeling errors.

Q4: Response Selection for Teacher Correction. (question 3)

A4: Thank you for your insightful comments. When all student rollouts in a group are incorrect, the teacher independently corrects each response using the ground truth formulation as a reference hint. We verify the corrections with the solver and randomly select one passing result as the SFT target. If none passes, we repeat correction independently, for at most three rounds in total. If all three fail, no SFT target is formed. With reference hints, approximately 90% of responses are corrected in the first round, and nearly all remaining responses within three rounds.

Best regards,

The authors.

References

[1] Dou, Shihan, et al. "CL-bench: A Benchmark for Context Learning." arXiv:2602.03587, 2026.

Reviewer_h4U7OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

MIND addresses two limitations of LLM post-training for automated optimization modeling: the scarcity of training examples targeting specific modeling errors and the sparsity of useful reinforcement-learning rewards on difficult problems. The paper is motivated by the observation that optimization modeling errors are frequently localized to individual semantic components, such as decision variables, constraints, and objective functions, while other components remain valid. Based on this insight, the authors introduce an error-driven reverse data synthesis strategy that constructs MIND-Train, a dataset emphasizing common and localized modeling failures. They also develop Dynamic Supervised Fine-Tuning Policy Optimization (DFPO), which supplements reinforcement learning with teacher-guided supervised correction when an entire rollout group fails. The evaluation includes six optimization modeling benchmarks and two open-source backbone models.

Critical strengths
  • The paper introduces a specific and potentially generalizable insight: many formulation errors can be corrected without replacing otherwise valid components. This provides a principled motivation for more targeted data synthesis and model correction.
  • The evaluation demonstrates substantial gains on two backbones. On Qwen2.5-7B, DFPO achieves 63.4% macro-average accuracy, compared with 61.5% for DAPO, 59.0% for SFT+GRPO, and 51.6% for SFT alone. These results support the value of dynamically integrating supervised correction into post-training.
Critical weaknesses
  • Objective-value agreement is insufficient to establish semantic correctness. Incorrect constraints, variable domains, or feasible regions may remain undetected, potentially rewarding mathematically invalid models.
  • DFPO depends on teacher-generated corrections, and the improvements may arise partly from injecting stronger-model knowledge rather than from the proposed localized error-learning mechanism.
  • The experiments establish improvements on established optimization-modeling benchmarks but provide limited evidence that localized error correction remains effective for strongly coupled formulations, distribution shifts, or unfamiliar optimization domains.
Questions for authors

The most significant technical limitation is the reliance on localized error decomposition and objective-based verification as approximations to semantic correctness. Although these assumptions may work for many standard optimization problems, they can fail when constraints are mutually dependent or when incorrect formulations happen to produce the correct optimal objective. DFPO also introduces dependence on the teacher's ability to identify and correct erroneous components without modifying valid ones, potentially propagating teacher mistakes into the training process. Experimentally, the reported gains do not fully disentangle the effects of specialized synthetic data, additional teacher supervision, dynamic activation, and reinforcement-learning reward design. Stronger matched-budget comparisons against always-on teacher distillation, conventional rejection-sampling fine-tuning, partial-response correction, and alternative dense-reward approaches would clarify whether DFPO itself provides a substantial advantage. Evaluation should also include constraint-level correctness, feasibility-region agreement, and held-out examples with intentionally coupled modeling errors.

Here are the key questions to be addressed. Q1: What proportion of optimization modeling errors are genuinely localized, and how does this proportion vary with problem complexity and the dependency structure of constraints? Q2: How often do incorrect formulations obtain the correct objective value, and how does this affect the accuracy of the proposed modeling-fidelity reward? Q3: How does DFPO perform with weaker teachers, different teacher architectures, or imperfect teacher corrections? Q4: Can the authors compare DFPO against ordinary teacher distillation using identical corrected outputs, training tokens, compute, and teacher-call budgets? Q5: Is the all-failed-rollout trigger optimal? How does performance vary with different activation thresholds or selectively corrected rollout groups? Q6: What is the independent contribution of MIND-Train relative to random synthetic examples, hard-example mining, and existing optimization-modeling datasets of the same size? Q7: Can the authors evaluate generated formulations using constraint-level correctness, feasible-region equivalence, variable-domain correctness, and adversarial cases where objective values coincide? Q8: Does MIND retain its advantage on unseen optimization problem families, highly coupled nonlinear formulations, or more difficult mixed-integer problems?

6 comments

Dear Reviewer h4U7,

We thank you for your detailed review and recognition of targeted correction and gains on both backbones.

Q1: Error Locality and Component Dependencies. (question 1)

A1: Thank you for this question. We distinguish error propagation during generation from error locality in completed formulations, and explain the evidence and its role in DFPO.

(1) Qualitative and quantitative evidence. Figure 1 provides a qualitative example in which an incorrect variable type leaves the parameters, objective function, and constraints valid. Figures 2 and 11 quantify the proportion of erroneous components in 100 failed responses per model. The mean component error ratios are 33%, 29%, 32%, and 28% for Qwen2.5-7B-Instruct, DeepSeek-V3, GPT-5, and Gemini-2.5-Pro, respectively, indicating substantial valid content in failed responses.

(2) Error propagation and final error locality. During generation, an incorrect variable definition may affect dependent constraints. Our analysis characterizes errors after the complete formulation has been produced. Several dependent components may be incorrect while other components remain valid. Thus, locality in the final formulation is compatible with error propagation during generation and does not require independent component errors.

(3) Error locality in DFPO. The teacher uses a reference formulation to correct the complete student response, allowing dependent variables and constraints to be revised together while retaining valid content. Corrections become SFT targets only after successful solver execution and objective value matching, and SFT learns the complete corrected response (Section 4.2 and Appendix B.4). Figures 16 and 17 illustrate local corrections on ComplexLP and OptMATH-Bench that preserve other valid components.

Q2: Formulation Correctness and the Modeling Fidelity Reward. (weakness 1 and questions 2 and 7)

A2: Thank you for raising these concerns. We use objective value matching as a practical evaluation proxy and the modeling fidelity reward as an auxiliary signal for error severity.

(1) Evaluation protocol and empirical support. Objective value matching is used in OptMATH [1], SIRL [7], and StepORLM [5]. Our evaluation requires successful solver execution and an absolute objective error below 10−610^{-6}. OptMATH [1] reports that 99.6% of cases with matching objective values in its manually inspected sample also have equivalent formulations.

(2) Controlled perturbation analysis. Following Appendix C.3, we apply controlled perturbations to 100 training formulations generated by LLMs with correct objective values, including 50 with fewer than 20 variables and 50 with more than 20. We perturb kk of NN constraints through constraint removal, right hand side (RHS) modification, coefficient perturbation, or inequality direction flipping, generating 100 variants per perturbation level. Infeasible, unbounded, and timed out candidates are excluded because a valid optimal objective comparison is unavailable.

For the within formulations analysis, we calculate Pearson and Spearman correlations between E=k/NE=k/N and mean absolute objective deviation across perturbation levels separately for each formulation, then average the respective coefficients across all 100 formulations. For the cross formulations analysis, we account for differences in objective scale using normalized deviation dnorm=∣Obj−Obj∗∣max⁡(∣Obj∣,∣Obj∗∣)d_{\mathrm{norm}}=\frac{\lvert\mathrm{Obj}-\mathrm{Obj}^{\ast}\rvert}{\max(\lvert\mathrm{Obj}\rvert,\lvert\mathrm{Obj}^{\ast}\rvert)}, where Obj∗\mathrm{Obj}^* is the reference objective. We combine the pairs of perturbation ratio and mean normalized deviation at each level from all 100 formulations into a single dataset and calculate its overall correlations:

AnalysisObjective deviation# FormulationsPearson rSpearman rho
Within formulationsAbsolute1000.7850.854
Cross formulationsNormalized1000.5660.603

These positive correlations indicate that objective value deviation increases with the proportion of perturbed constraints on average, supporting its use as an auxiliary signal for modeling errors.

(3) Training benefit and reward weighting. We use R=αRfid+(1−α)RaccR=\alpha R_{\mathrm{fid}}+(1-\alpha)R_{\mathrm{acc}} with α=0.2\alpha=0.2, giving the fidelity and accuracy terms a weight ratio of 1:4. Figure 13 and Appendix C.6 compare Qwen2.5-7B-Instruct DAPO runs trained on the same 10K instances for seven epochs. Compared with accuracy reward alone, adding the fidelity term improves greedy Pass@1 on OptMATH-Bench from 9.0% to 17.5% and ComplexLP from 39.4% to 44.3%. Increasing α\alpha to 0.6 changes the ratio to 3:2 and reduces performance on four of six benchmarks relative to accuracy reward alone. This is consistent with noise in the imperfect fidelity proxy becoming more influential at higher weight and supports retaining accuracy as the primary objective. Infeasible or unbounded student formulations are excluded from reward computation and receive neither reward.

(4) Relative reward feedback. OptMATH reports that 0.4% of manually inspected cases had matching objective values but nonequivalent formulations [1]. An incorrect formulation may occasionally match the reference objective and receive a positive accuracy reward. When the model applies the same erroneous modeling pattern to another problem, it receives no accuracy reward if the resulting objective fails the reference check. If its reward falls below the group average, GRPO [8] assigns a negative relative advantage, providing a training signal that discourages that response. This mechanism can mitigate the reinforcement of erroneous responses caused by occasional false positive rewards.

Q3: Teacher Sensitivity and Correction Reliability. (weakness 2 and question 3)

A3: We appreciate this suggestion. Reference guidance enables useful correction even when unaided accuracy is low. Qwen3.8-27B supports nonthinking (nothink) and thinking modes, with low, medium, and extra high (xhigh) reasoning effort in thinking mode. We first evaluate correction using reference information in nonthinking mode and at each reasoning effort level in thinking mode. We then conduct a separate training comparison to test whether a teacher model using the same backbone can improve the student.

(1) Correction with reference context. The correction test uses 82 of 166 OptMATH-Bench questions with correct full reference responses from GPT-6-astra. For each mode and reasoning effort level, Qwen3.8-27B first answers each original problem directly without a hint, using the paper's evaluation prompt. For an incorrect initial response, correction follows the paper's teacher correction prompts in two stages, both keeping the mode and reasoning effort fixed. First, the model receives the original problem and the full GPT reference response and generates a complete PySCIPOpt reference solution. Second, it receives the original problem, its own initial response, and the reference solution it generated in the first stage, then locally repairs the initial response and outputs a complete corrected response. Initially correct responses are retained, and only incorrect responses undergo repair. There is no iterative solver feedback.

Direct response accuracy and accuracy after correction both use all 82 questions as the denominator and the same PySCIPOpt execution and objective value matching criterion. Temperature is 1.0 and the maximum generation length is 73,728 tokens.

ModelReasoning modeReasoning effortDirect response accuracyAccuracy after correctionGain (pp)Mean direct tokensMean additional correction tokensWord level F1 scoreEmbedding cosine similarity
Qwen3.8-27BNonthinkingN/A2/82 (2.44%)29/82 (35.37%)+32.936,113.0118,300.800.62280.9291
Qwen3.8-27BThinkingLow15/82 (18.29%)65/82 (79.27%)+60.986,384.8920,041.390.97240.9934
Qwen3.8-27BThinkingMedium17/82 (20.73%)68/82 (82.93%)+62.206,740.3421,452.150.94230.9901
Qwen3.8-27BThinkingExtra high23/82 (28.05%)66/82 (80.49%)+52.4427,532.3250,539.720.94830.9854

With medium reasoning effort, correction using reference context increases accuracy from 20.73% to 82.93%. This use of reference context is consistent with context learning [4]. Word level F1 score and embedding cosine similarity compare each initial response with its corrected response and are averaged only over successful corrections from incorrect to correct. The scores indicate substantial textual retention. Reasoning effort is not applicable in nonthinking mode. Gain is measured in percentage points (pp). Tokens count API input and output, including reasoning tokens. Additional correction tokens cover both reference generation and repair, excluding direct answering.

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

This paper tackles automated optimization modeling—translating natural language problem descriptions into executable mathematical formulations—and identifies two bottlenecks: the high cost of generating high-quality training data and sparse reward signals in RL-based post-training, especially on difficult problems. The authors' key empirical observation is that LLM errors in this domain are localized: when a formulation is wrong, the model typically errs in only a small subset of components rather than producing an entirely incorrect model, in contrast to general mathematical reasoning where a single early error cascades through all subsequent steps. Motivated by this, they propose MIND, a two-stage framework. The first stage is an error-driven reverse data synthesis pipeline that avoids costly expert-annotated seeds by sampling existing optimization datasets, eliciting failures from a base model, having a strong LLM identify and extract the error patterns, and then synthesizing new, deliberately harder questions that embed those traps—singly or in combination. Quality is enforced through solver executability checks plus bidirectional validation, where another LLM re-solves the synthesized problem and its optimal value is compared against ground truth. The second stage, Dynamic Supervised Fine-Tuning Policy Optimization (DFPO), addresses sparse rewards by combining a graded fidelity proxy—normalized objective-value deviation, justified by controlled experiments showing modeling error correlates with objective deviation—with objective accuracy. Crucially, when an entire rollout group fails, DFPO dynamically activates an SFT branch in which a teacher model corrects the failed student responses, preserving valid portions and fixing only localizable errors, so RL and SFT losses are jointly optimized. Experiments across several benchmarks and two backbones show substantial improvements over base models and comparable-scale baselines, with ablations confirming the value of both the synthesized data and DFPO, including the contribution of teacher-guided response correction.

Critical strengths

1.Genuinely insightful core observation. The localized-error finding (Fig. 1, 2, 11) is simple, well-quantified across four model families, and directly motivates both the data pipeline and the correction mechanism. The framing contrast between math reasoning (one early error cascades) vs. optimization modeling (errors confined to components) is pedagogically clear.

2.Coherent "observation → data → training" story. The two stages are not bolted together: the same localization property justifies (a) synthesizing problems with embedded error traps and (b) teacher-correcting only the wrong portions of failed rollouts. This is a tighter narrative than many "data + RL" papers.

3.Cheap, scalable data synthesis. Avoiding costly expert-annotated seed data by reusing existing datasets and inverting them into harder problems is practical. The two-stage quality control (solver executability + bidirectional LLM re-solve validation) is a sensible guard against noisy synthetic labels, and the single- vs. multi-error ablation (MIND-Mix > Single > Multi) is a useful empirical finding—multi-error data is too hard to learn from directly.

4.DFPO targets a real failure mode. Providing supervised signal exactly for all-failed rollout groups is well-motivated (OptMATH has only 2.3% Avg@8, so RL alone gets almost no gradient), and the ablation isolating response correction (61.9% vs. 63.4%) supports the design choice. The graded fidelity reward is also empirically grounded

Critical weaknesses

1.Circularity risk in the core hypothesis. The reward's fidelity term R_fid is a proxy for formulation correctness, justified by a correlation between error ratio and objective deviation (Eq. 2, C.3). But this correlation is established on synthetically perturbed constraints of five small problems (6–20 vars), not on real LLM formulation errors. The assumption that graded objective deviation is a faithful fidelity proxy in the wild is plausible but not directly validated. Also, α sensitivity shows α=0.6 hurts—so the fidelity signal is a marginal auxiliary, and the paper doesn't fully explain why the correlation holds well enough to be useful.

2.Teacher dependence and cost. DFPO's gains rely on DeepSeek-V3 corrections, and the paper itself admits (C.11) that supervision quality depends on reference-solution quality—an unresolved practical bottleneck. Teacher usage during training is non-trivial (6,356 tokens/instance on 100 samples; routing starts at 65.5%), and the "Pro" experiment uses DeepSeek-V4-Pro for synthesis. The method's cost profile versus plain DAPO is not reported, so scalability claims are hard to assess.

3.Modest/uneven gains and one regression. On OptiBench, MIND-Qwen2.5-7B gains only marginally (56.7% vs. 55.9% SFT), attributed to tabular data underrepresentation—but this is a coverage gap of the synthesis pipeline that isn't fixed. In Table 8, MIND-3K actually drops on OptMATH (7.8% vs. 10.8% for OR-Instruct-3K), and the main-method OptMATH score (36.7%) is driven by DFPO, not data. The data-synthesis benefit is thus benchmark-dependent and not cleanly isolated.

Questions for authors

1.Reward validity: The fidelity reward is justified by correlations measured on synthetically perturbed constraints (C.3). Have you validated that normalized objective deviation remains a monotone proxy for formulation fidelity on real LLM-generated formulations (e.g., by correlating R_fid with expert-annotated E on the 300-sample analysis in C.8)? If the correlation is weaker in the wild, does DFPO still help?

2.Cost accounting: What is the wall-clock/GPU-hour and teacher-token overhead of DFPO relative to DAPO and SFT+GRPO at matched final accuracy? Given routing starts at 65.5%, how does cost scale with model capability (i.e., for a stronger base model that fails less often)?

3.Fair baseline comparison: Why does SFT use LoRA (rank 8) while RL methods are full-parameter? Would a full-parameter SFT baseline (matched epochs/data) close part of the gap? Similarly, how does DFPO compare to rejection-sampling SFT using only successful rollouts, or to DAPO trained longer on the same 10K instances?

6 comments

Dear Reviewer XnpC,

Thank you for your thoughtful review. We are encouraged by your recognition of our motivation based on localized errors, targeted correction, and scalable synthesis.

Q1: Validity of the Modeling Fidelity Reward. (weakness 1 and question 1)

A1: Thank you for this question. We use the modeling fidelity reward as an auxiliary signal for error severity alongside the binary accuracy reward. Dynamic SFT is triggered when all student rollouts in a group have zero binary accuracy reward, allowing these groups to receive corrected SFT targets verified by the solver even when the fidelity signal provides limited guidance.

(1) Training benefit and reward weighting. Figure 13 and Appendix C.6 compare Qwen2.5-7B-Instruct DAPO runs trained on the same 10K instances for seven epochs with R=αRfid+(1−α)RaccR=\alpha R_{\mathrm{fid}}+(1-\alpha)R_{\mathrm{acc}}. At α=0.2\alpha=0.2, the fidelity and accuracy weights have a ratio of 1:4. Compared with accuracy reward alone (α=0\alpha=0), this improves greedy Pass@1 on OptMATH-Bench from 9.0% to 17.5% and ComplexLP from 39.4% to 44.3%. Increasing α\alpha to 0.6 changes the weight ratio to 3:2 and reduces performance on four of six benchmarks relative to α=0\alpha=0. This pattern is consistent with noise in the imperfect fidelity proxy becoming more influential at higher weight. The gains at α=0.2\alpha=0.2 support using the modeling fidelity reward as an auxiliary signal for error severity, with accuracy remaining the primary objective.

(2) Controlled perturbation analysis. Following Appendix C.3, we apply controlled perturbations to 100 training formulations generated by LLMs with correct objective values, including 50 with fewer than 20 variables and 50 with more than 20. We perturb kk of NN constraints through constraint removal, right hand side (RHS) modification, coefficient perturbation, or inequality direction flipping, generating 100 variants per perturbation level. Candidates that are infeasible, unbounded, or time out are excluded because a valid optimal objective comparison is unavailable. Within each formulation, we calculate Pearson and Spearman correlations between the perturbation ratio E=k/NE=k/N and mean absolute objective deviation across perturbation levels, then average the respective coefficients across formulations:

Formulation cohort# FormulationsPearson rSpearman rho
Variables <20500.9020.944
Variables >20500.6690.765
All formulations1000.7850.854

The within formulations analysis calculates correlations separately for each formulation and then averages them. For the cross formulations analysis, we first account for differences in objective scale using normalized deviation dnorm=∣Obj−Obj∗∣max⁡(∣Obj∣,∣Obj∗∣)d_{\mathrm{norm}}=\frac{\lvert\mathrm{Obj}-\mathrm{Obj}^{\ast}\rvert}{\max(\lvert\mathrm{Obj}\rvert,\lvert\mathrm{Obj}^{\ast}\rvert)}, where Obj∗\mathrm{Obj}^* is the reference objective. We then combine the pairs of perturbation ratio and mean normalized deviation at each perturbation level from all 100 formulations into a single dataset and calculate its overall Pearson and Spearman correlations:

AnalysisObjective deviation# FormulationsPearson rSpearman rho
Within formulationsAbsolute1000.7850.854
Cross formulationsNormalized1000.5660.603

The positive correlations within and across formulations indicate that objective value deviation increases with the proportion of perturbed constraints on average, supporting its use as an auxiliary signal for modeling errors.

Q2: Teacher Dependence and Computational Cost. (weakness 2 and question 2)

A2: Thank you for this suggestion. DFPO supplies supervision for difficult prompts without successful rollouts, achieves gains that persist after additional baseline training, and can use the same backbone in thinking mode with medium reasoning effort for teacher correction.

(1) Learning signals for difficult prompts. When all student rollouts in a group are incorrect, the binary accuracy reward provides no contrast for GRPO updates. DAPO filters out these groups, whereas DFPO uses teacher corrections as supervised targets through dynamic SFT. As training progresses, the proportion of prompts requiring correction falls from 65.5% to 18.3% (Appendix C.1). A stronger student therefore requires fewer teacher calls on a fixed problem set, while harder problems can still produce groups with no successful rollout.

(2) Training costs and extended training. For Qwen2.5-7B-Instruct, we report cumulative student A100 GPU hours, Avg@8 on each of the six benchmarks, and their macro average. Entries show mean accuracy (%) ± sample standard deviation (percentage points) across eight inference sampling seeds with fixed checkpoints (temperature = 1.0, top_p = 0.7). Rows marked (+5 epochs) continue the trained checkpoint for five additional epochs on the same 10K instances.

MethodStudent A100 GPU hoursNL4OptIndustryOREasyLPComplexLPOptMATH-BenchOptiBenchMacro AVG
DFPO78096.0 ± 0.6136.9 ± 2.3091.5 ± 0.2561.9 ± 1.2034.9 ± 2.0956.6 ± 0.4863.0 ± 0.64
DAPO71896.1 ± 0.7835.4 ± 2.0791.7 ± 0.4858.5 ± 1.2826.0 ± 1.0456.7 ± 0.3860.7 ± 0.38
SFT+GRPO65294.0 ± 1.0031.3 ± 2.1290.1 ± 0.3253.6 ± 2.6929.9 ± 1.6756.4 ± 0.5059.2 ± 0.48
DAPO (+5 epochs)85696.5 ± 0.3134.0 ± 2.5191.7 ± 0.3359.4 ± 1.2126.4 ± 0.9456.9 ± 0.7260.8 ± 0.57
SFT+GRPO (+5 epochs)77294.2 ± 0.7735.1 ± 3.0090.0 ± 0.4855.9 ± 1.0929.5 ± 1.2056.6 ± 0.7560.2 ± 0.77

After five additional epochs, DAPO and SFT+GRPO still fall short of DFPO in macro average Avg@8 despite using greater or comparable student GPU budgets. This supports the value of directly supplying supervision to groups without successful rollouts.

(3) Correction with the same backbone. Qwen3.8-27B supports nonthinking (nothink) and thinking modes, with low, medium, and extra high (xhigh) reasoning effort in thinking mode. We first examine whether reference context supports response correction in both modes and across reasoning effort levels. On the same 82 of 166 OptMATH-Bench questions answered correctly by GPT-6-astra, the model independently answers without hints using the paper's evaluation prompt. For each wrong response, the model follows the teacher correction prompts with its mode and reasoning effort held fixed: it first generates a PySCIPOpt reference solution from the question and GPT-6-astra's full response, then repairs the erroneous parts of its initial response using the generated reference and outputs a complete corrected response.

Temperature is 1.0 and the maximum generation length is 73,728 tokens. Both accuracies use all 82 questions and the same PySCIPOpt execution and objective value matching criterion. Similarity is computed between each wrong response and its corresponding corrected response, then averaged over successful corrections.

Reasoning effort is not applicable in nonthinking mode. Gain is measured in percentage points (pp). Tokens count API input and output, including reasoning tokens. Additional correction tokens cover both reference generation and repair, excluding direct answering.

ModelReasoning modeReasoning effortDirect response accuracyAccuracy after correctionGain (pp)Mean direct tokensMean additional correction tokensWord level F1 scoreEmbedding cosine similarity
Qwen3.8-27BNonthinkingN/A2/82 (2.44%)29/82 (35.37%)+32.936,113.0118,300.800.62280.9291
Qwen3.8-27BThinkingLow15/82 (18.29%)65/82 (79.27%)+60.986,384.8920,041.390.97240.9934
Qwen3.8-27BThinkingMedium17/82 (20.73%)68/82 (82.93%)+62.206,740.3421,452.150.94230.9901
Qwen3.8-27BThinkingExtra high23/82 (28.05%)66/82 (80.49%)+52.4427,532.3250,539.720.94830.9854

Correction with reference context improves accuracy over direct answering in nonthinking mode and at all three reasoning effort levels in thinking mode, while successful corrections retain substantial student content. This use of supplied context is consistent with the context learning perspective in CL-bench [3].

To test whether this correction capability provides useful training supervision, we train DAPO and DFPO from the same Qwen3.8-27B base model on the same 10K instances for two epochs. Both students use nonthinking mode, and DFPO uses the same backbone in thinking mode with medium reasoning effort as its teacher, with reference formulations as hints. Both students are evaluated in nonthinking mode under identical Pass@1 settings. We report Pass@1 accuracy (%) on each of the eight benchmarks, including ComplexOR (18 instances) [1] and MIPLIB-NL (220 valid instances) [2], together with their macro average:

ModelNL4OptIndustryOREasyLPComplexLPOptMATH-BenchOptiBenchComplexORMIPLIB-NLMacro AVG
Qwen3.8-27B (DAPO)86.156.092.754.749.454.455.629.159.7
Qwen3.8-27B (DFPO)89.855.092.862.155.454.950.036.462.0

DFPO improves on six of eight benchmarks. These results show that thinking mode with medium reasoning effort can provide effective correction supervision for the nonthinking mode of the same backbone, addressing dependence on a separate stronger teacher model.

Reviewer_U8N2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MIND trains large language models (LLMs) to turn optimization problems in natural language into solver code. The basis is a local-error result: in 100 failed responses, the mean error ratio (the fraction of wrong components) is 0.33. MIND synthesizes problems around base-model errors, and DFPO (Dynamic Supervised Fine-Tuning Policy Optimization) adds a supervised loss on teacher-corrected responses when all samples fail. On six benchmarks, the macro-average accuracy rises from 48.9% to 63.4% for Qwen2.5-7B-Instruct and from 37.8% to 70.0% for Qwen3-8B.

Critical strengths
  • The local-error result holds for four LLMs, with mean error ratios of 0.28 to 0.33. DFPO uses it: the teacher corrects the failed student response and keeps its correct parts, unlike the ground-truth answer or hint remedies in Sec. 4.2. In the "DFPO w/o correction" ablation, OptMATH accuracy falls from 36.7% to 29.5%.
  • MIND-Qwen3-8B reaches 70.0% macro-average accuracy, above SIRL-Qwen2.5-32B (69.2%) and equal to Claude-Opus-4.7. The results also hold with the Gurobi solver and under random sampling.
Critical weaknesses
  • In DAPO training runs, MIND-3K beats its seed set OR-Instruct-Data-3K by only 0.9 points and loses 3.0 points on OptMATH. No control uses the same synthesis LLM and filters without error patterns, so the source of this gain is unclear. Also, MIND-Train gives only 5,000 of the 10,000 main training instances, and no main-size run uses seed data only. So Table 1 does not isolate the value of MIND-Train.
  • DFPO beats DAPO by 1.9 points and "DFPO w/o correction", which uses the same teacher, by 1.5 points. Most of the gain is on OptMATH-Bench (166 questions). On the other five benchmarks, the mean gain is 0.4 points or less, and the paper reports no seeds or variance. The paper also reports no train/test overlap check, although OptMATH-Train, from the same source as OptMATH-Bench, gives 4,000 training instances.
Questions for authors
  • Did you check exact and near-duplicate overlap between all training data and the six benchmarks?
  • Can you report the mean and standard deviation over three or more seeds for each Table 2 row?
6 comments

Dear Reviewer U8N2,

Thank you for your careful review and constructive comments.

Q1: Contribution of Reverse Data Synthesis. (weakness 1)

A1: Thank you for this comment. We agree that Table 1 evaluates the full framework on a mixture of MIND-Train and seed data and does not isolate the contribution of MIND-Train. We clarify below our control of training data size, analysis of synthesis strategies, and evidence for scalability.

(1) Training data controls. We use 10K training instances to match SIRL's training set size. Our separate comparison in RQ3 controls dataset size, backbone, and training algorithm: training Qwen2.5-7B-Instruct with DAPO on MIND-3K or OR-Instruct-Data-3K yields macro average greedy Pass@1 of 53.2% and 52.3%, respectively, across six benchmarks (Table 8). This supports the competitive training utility of the synthesized data.

(2) Analysis of error patterns. Appendix C.5 further examines how error patterns affect data difficulty and training utility. The initial Qwen2.5-7B-Instruct model achieves 41.2% accuracy on examples synthesized from multiple error patterns and 52.9% on those synthesized from single error patterns. After seven DAPO epochs, macro average greedy Pass@1 across six benchmarks is 53.2% for Mix-3K, 52.6% for Single-1.5K, and 49.9% for Multi-1.5K. The mixed set performs best among these configurations, while training only on the more challenging examples is less effective. This motivates including examples at different difficulty levels. Figures 6 and 7 illustrate how individual and combined error patterns guide problem synthesis.

(3) Scalability and data release. Our contribution is a scalable synthesis method that complements existing approaches through observed modeling error patterns. Reusing public datasets as seeds reduces the need for newly annotated seed problems, while code validation and bidirectional validation provide quality control (Section 4.1). The Pro experiments in Appendix D show that continuing the existing MIND checkpoints for one DFPO epoch on 50K instances sampled from 120K additional examples synthesized from OptMATH-Train further improves macro average Avg@8. We will release approximately 130K MIND-Train instances as a training resource for the automated optimization modeling community.

Q2: DFPO's Gains on OptMATH-Bench. (weakness 2)

A2: Thank you for this observation. We address it through the learning mechanism and two additional evaluations.

(1) Sparse rewards on difficult problems. To understand the larger gains on OptMATH-Bench, we compare the initial Qwen2.5-7B-Instruct model across six benchmarks (Table 1 and Appendix C.13). We report Pass@1 under greedy decoding (temperature = 0.0), and Avg@8 and Pass@8 from eight responses per question (temperature = 1.0, top_p = 0.7). Avg@8 averages accuracy over the eight responses, while Pass@8 measures the fraction of questions with at least one correct response. All values are percentages.

MetricNL4OptIndustryOREasyLPComplexLPOptMATH-BenchOptiBenchMacro AVG
Pass@189.027.089.431.53.053.248.9
Avg@870.720.079.021.12.345.539.8
Pass@897.641.096.652.215.165.361.3

OptMATH-Bench has the lowest Avg@8 and Pass@8, indicating that correct rollouts are particularly difficult to obtain. This highlights the sparse reward challenge for RLVR on training problems of comparable difficulty. When all responses in a group are incorrect, the binary accuracy reward provides no contrast for RL updates. DAPO filters out these groups, whereas DFPO uses corrected student responses as targets for dynamic SFT. This supervision supplies learning signals for difficult prompts and helps explain DFPO's larger gains on OptMATH-Bench.

(2) Evaluation on challenging benchmarks. We directly evaluate our existing Qwen2.5-7B and Qwen3-8B checkpoints, trained on textual problem descriptions, on ComplexOR (18 instances) [1] and MIPLIB-NL (220 instances) [2] without additional training. For ComplexOR, we use the corrected labels from the LLM4OR survey [3]. MIPLIB-NL targets industrial optimization modeling, where problems can involve 10310^3 to 10610^6 or more variables and constraints. It separates model descriptions from instance data, requiring generated code to load external files. Three of the original 223 MIPLIB-NL instances have empty data files and cannot be instantiated, so we exclude them. Both benchmarks use Pass@1 accuracy (%) based on solver execution and objective value matching:

ModelComplexORMIPLIB-NL
SIRL-Qwen2.5-7B†22.20.5
SIRL-Qwen2.5-32B†44.41.4
OptMATH-Qwen2.5-7B†38.90.0
OptMATH-Qwen2.5-32B†27.85.5
DAPO-Qwen2.5-7B (PySCIPOpt)44.45.0
DFPO w/o correction-Qwen2.5-7B (PySCIPOpt)50.04.5
MIND-Qwen2.5-7B (PySCIPOpt)50.07.3
MIND-Qwen2.5-7B (Gurobi)66.77.7
MIND-Qwen3-8B (PySCIPOpt)55.623.6

† Results reported in [2], which excludes infeasible and open MIPLIB-NL instances. Our results use 220 instances. MIND models are trained with DFPO.

With Qwen2.5-7B and PySCIPOpt, DFPO improves over DAPO by 5.6 and 2.3 percentage points on ComplexOR and MIPLIB-NL, respectively. It matches DFPO w/o correction on ComplexOR and improves MIPLIB-NL accuracy from 4.5% to 7.3%, extending the observed benefits beyond OptMATH-Bench.

(3) Evaluation with a larger backbone as both student and teacher. Qwen3.8-27B supports nonthinking mode and thinking mode with low, medium, or extra high reasoning effort.

We compare direct answering and correction on the 82 of 166 OptMATH-Bench questions answered correctly by GPT-6-astra. For each mode and reasoning effort, Qwen3.8-27B first answers without hints using the paper's evaluation prompt. For each wrong response, it follows the paper's teacher correction prompts to generate a PySCIPOpt reference solution from the question and GPT-6-astra's full response, then repairs its own wrong response using that reference. The mode and reasoning effort remain fixed throughout.

All generations use temperature 1.0 and a maximum length of 73,728 tokens. Both accuracies use all 82 questions and the same PySCIPOpt execution and objective value matching criterion. Similarity scores compare wrong responses with their corresponding corrected responses, averaged over successful corrections. Gain is in percentage points (pp). Tokens include API input and output, including reasoning tokens. Additional correction tokens cover both correction stages, excluding direct answering.

ModelReasoning modeReasoning effortDirect response accuracyAccuracy after correctionGain (pp)Mean direct tokensMean additional correction tokensWord level F1 scoreEmbedding cosine similarity
Qwen3.8-27BNonthinkingN/A2/82 (2.44%)29/82 (35.37%)+32.936,113.0118,300.800.62280.9291
Qwen3.8-27BThinkingLow15/82 (18.29%)65/82 (79.27%)+60.986,384.8920,041.390.97240.9934
Qwen3.8-27BThinkingMedium17/82 (20.73%)68/82 (82.93%)+62.206,740.3421,452.150.94230.9901
Qwen3.8-27BThinkingExtra high23/82 (28.05%)66/82 (80.49%)+52.4427,532.3250,539.720.94830.9854

Correction improves accuracy in all four settings, with successful repairs retaining substantial initial content. This use of reference context is consistent with context learning in CL-bench [4].

To evaluate the training benefit, we train DAPO and DFPO from the same Qwen3.8-27B base model on the same 10K instances for two epochs. Both students use nonthinking mode. DFPO uses the same backbone in thinking mode with medium reasoning effort as its teacher, with training reference formulations as hints. Both students are evaluated in nonthinking mode under identical Pass@1 settings across eight benchmarks (%):

ModelNL4OptIndustryOREasyLPComplexLPOptMATH-BenchOptiBenchComplexORMIPLIB-NLMacro AVG
Qwen3.8-27B (DAPO)86.156.092.754.749.454.455.629.159.7
Qwen3.8-27B (DFPO)89.855.092.862.155.454.950.036.462.0

DFPO improves on six of eight benchmarks, including gains of 7.4 percentage points on ComplexLP and 7.3 on MIPLIB-NL, extending its benefits beyond OptMATH-Bench.

Q3: Training and Benchmark Overlap. (weakness 2 and question 1)

A3: Thank you for your insightful suggestions. We address your questions below.

(1) Audit procedure. We audited all 10K main training questions against 1,961 questions from the six benchmarks in three steps.

Step 1. Exact matching. We compared the original question texts and repeated the comparison after Unicode and whitespace normalization. Both checks found zero exact text duplicates.

Step 2. Similarity retrieval. For each benchmark question, we retrieved the top five similar training questions using TF-IDF representations at the word and character levels, combining the two rankings through reciprocal rank fusion. This produced 9,805 candidate pairs.

Step 3. LLM classification. We screened the retrieved pairs and used GPT-6-astra to review flagged matches and classify their relationships as strict same instances, scenario variants, parameter variants, structural variants, or no specific relationship.

Comments on the paper

7 comments
AuthorsOpenReview ↗

Global Response

Dear Area Chair and Reviewers,

We sincerely apologize for the delay in our response. We greatly appreciate the time and effort that the Area Chair and each reviewer devoted to providing valuable feedback. We have taken every comment seriously, but completing the additional experiments prevented us from responding earlier. We summarize the main points of our rebuttal below.

1. Competitive and scalable reverse data synthesis. Our method generates new problems and solutions from observed modeling errors, providing an automated, scalable complement to existing synthesis methods. Equal size comparisons show competitive training utility, and training on the expanded dataset further improves existing MIND models. We will release approximately 130K MIND-Train instances for the automated optimization modeling community.

2. Learning signals for difficult problems. DFPO addresses sparse rewards by providing corrected responses through dynamic SFT when all student rollouts in a group are incorrect. The initial Qwen2.5-7B-Instruct model's low Pass@8 on OptMATH-Bench supports this motivation. Additional training of DAPO and SFT+GRPO does not close the observed performance gap, supporting the value of correction supervision beyond longer training.

3. Gains on more challenging benchmarks. We evaluate existing Qwen2.5-7B and Qwen3-8B checkpoints without additional training on ComplexOR (18 questions) [1] and MIPLIB-NL (220 instances after excluding three with empty data files) [2]. With Qwen2.5-7B, DFPO outperforms DAPO on both benchmarks, extending the observed gains to industrial optimization problems in MIPLIB-NL that require generated code to load external instance data.

4. Correction within the same backbone. Qwen3.8-27B supports both nonthinking and thinking modes. Our correction experiments show that reference context enables effective response repair, consistent with context learning [3]. Using the same backbone in thinking mode with medium reasoning effort as the teacher, DFPO trains a nonthinking student and outperforms DAPO on six of eight benchmarks under matched data and training duration.

Detailed results and methodological clarifications are provided in our individual reviewer responses.

Best regards,

The authors.

References

[1] Xiao, Ziyang, et al. "Chain-of-Experts: When LLMs Meet Complex Operations Research Problems." ICLR, 2024.

[2] Li, Zhong, et al. "Constructing Industrial-Scale Optimization Modeling Benchmark." ICML, 2026.

[3] Dou, Shihan, et al. "CL-bench: A Benchmark for Context Learning." arXiv:2602.03587, 2026.

Dear Reviewers,

Please note that the author rebuttal is now available. We kindly ask you to read it, join the discussion, and update the review as soon as possible. Your prompt action will help us proceed to the final decision. Thanks.

Best regards

AC

Dear Reviewer h4U7,

Following your suggestions, we expanded the main text to ten pages with additional experiments in Section 6 and a discussion of limitations in Section 7, and added the MIPLIB-NL formulation audit to Appendix C.15. Below, we address your remaining questions in four groups covering data synthesis and quality, post-training, the evaluation protocol, and experimental results.

1. Data Synthesis and Data Quality

Q8: Have you compared error-driven synthesis with random synthesis and hard-example mining at the same dataset size?

A8: Error-driven reverse data synthesis uses observed modeling errors to generate new problems and solutions from public seed data. Reusing these seeds reduces the need for expert annotation and enabled us to generate 120K additional instances (Appendix D). With Qwen2.5-7B-Instruct, MIND-3K performed competitively against its seed dataset, OR-Instruct-Data-3K, the only size-matched control in this comparison (Table 13). We plan to release approximately 130K MIND-Train instances as a public training resource for the automated optimization modeling community. A team of senior researchers in operations research issued an open call for contributions to an optimization modeling benchmark they were developing for a paper. We used GPT-5.6-sol to select two synthesized MIND-Train instances with relatively large numbers of decision variables and submitted them anonymously. After review by external operations research experts, both instances were accepted for inclusion in the benchmark. These results suggest that error-driven reverse data synthesis complements existing synthesis methods and that MIND-Train is a useful training resource for the automated optimization modeling community.

2. Post-Training

Q9: Does the same-backbone experiment separate the effect of correcting a student response from the teacher's knowledge, access to a reference, and reasoning effort?

A9: On the 82 of 166 OptMATH-Bench questions answered correctly by GPT-6-astra, increasing Qwen3.8-27B's reasoning effort from medium to extra high raised its unaided accuracy from 17/82 (20.73%) to 23/82 (28.05%). At fixed medium effort, reference-guided correction raised accuracy from 17/82 (20.73%) to 68/82 (82.93%). Qwen revised each incorrect response using DFPO's two-stage correction prompts with GPT-6-astra's solution as a reference hint (see our first-round response to Q3, part 1). This use of supplied context is consistent with context learning [1]. We believe this improvement depends more on the teacher's context learning ability than on its optimization modeling ability. On these difficult questions, using reference context to correct answers at medium effort yielded a larger accuracy gain (62.2 percentage points) than increasing reasoning effort alone (7.3 points).

Q10: Does the evidence show how error locality varies with complexity or constraint coupling, and whether correction preserves valid components?

A10: Our locality observation concerns the completed formulation. During generation, errors may propagate across dependent components. For example, an incorrect variable definition may affect constraints that use that variable. Figure 1 illustrates error locality in one completed formulation. The variable type is incorrect while the objective and constraints remain valid. Figures 2 and 11 quantify the outcome across 100 failed responses from each of four LLMs. The mean fraction of incorrect components ranged from 28% to 33% across models.

This observation motivates both parts of MIND. Reverse data synthesis uses observed error patterns to generate new problems and solutions. When all student rollouts for a prompt are incorrect, DFPO uses the completed responses as the starting point for correction. Guided by a reference formulation, the teacher can revise dependent components together and retain valid ones. A correction that passes solver execution and objective matching can provide an SFT target for a group whose binary rewards are all zero.

Appendix C.9 illustrates what was preserved. In the diet problem, the variable domains changed while the cost objective and dietary constraints remained intact. In the scheduling problem, the teacher corrected the machine capacity and makespan constraints while preserving the precedence constraints. Our analysis measures the fraction of incorrect components in the completed formulation after any error propagation during generation, and DFPO constructs a verified SFT target by correcting the affected parts while retaining valid content.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.