acceptodds
Under review as a conference paper at ICLR 2027

RELIVE-OR: Retrospective Learning through Integrated Validation of Experience for Operation Research Modelling

Abstract

Synthetic data is increasingly used to fine-tune large language models (LLMs) for operations research (OR) modeling, yet scaling synthesis alone does not necessarily produce better OR modelers. We view each problem–program pair as a modeling experience and OR training as a process of acquiring, validating, revisiting, and reusing such experiences. This perspective exposes a key limitation of existing synthesis pipelines: beyond semantic, code-structural, and outcome consistency, a synthetic experience should remain consistent with the OR model specification represented by its trusted seed. Otherwise, it may remain executable and numerically valid while becoming misspecified with respect to the intended modeling principle. We propose RELIVE-OR—REtrospective Learning through Integrated Validation of Experience for OR modeling—with the objective of maximizing structural freedom while preserving OR model specification. Starting from a small trusted seed set, RELIVE-OR synthesizes structurally diverse experiences directly in executable code space and applies closed-loop verification and repair for semantic, code-structural, and outcome consistency. The verified experiences are first used for broad consolidation to expose the learner to diverse OR structures. RELIVE-OR then retrospectively estimates the influence of each verified experience on a trusted seed objective and re-selects experiences that remain aligned with the intended OR model specification for focused consolidation. Finally, the verified experience bank is reused at test time as a non-parametric reference memory for gradient-guided selection among executable candidates, without additional parameter updates. Across six established OR modeling benchmarks, RELIVE-OR enables an 8B model to achieve state-of-the-art or competitive performance and perform on par with substantially larger frontier models.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Operation Research Foundation Model, Code Generation, OR model specification

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_LQGTOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RELIVE-OR, a framework for improving LLM-based operations research modeling through synthetic experience learning. It combines structural data synthesis, closed-loop verification, and influence-based sample selection for fine-tuning, while introducing FREE for update-free test-time candidate selection. Experiments across six OR modeling benchmarks demonstrate competitive performance with an 8B model.

Critical strengths
  1. Meaningful research problem. The paper addresses an important challenge in LLM-based OR modeling: generating structurally diverse training data while maintaining modeling correctness.

  2. Comprehensive framework and competitive results. The paper presents an integrated framework covering data synthesis, validation, training, and inference. Experiments across six OR benchmarks demonstrate competitive performance with an 8B model.

Critical weaknesses
  1. Several key aspects of the proposed method remain unclear. Although Equations (8)–(9) define seed-relative OR model-specification preservation, the paper does not sufficiently explain how these specifications are constructed and verified in practice. Similarly, Section 2.1.3 introduces structural consistency checks, but their actual implementation is not clearly described. These missing details make it difficult to assess the reliability and reproducibility of the validation framework.

  2. The framework combines several existing techniques, such as influence-based data selection, repeated SFT, and reference-guided candidate selection. However, the connection between these components and the paper's central motivation of preserving OR model specifications is not sufficiently clear. Moreover, the experiments do not adequately isolate the contributions of individual components, making it difficult to determine whether the improvements arise from the proposed mechanisms or from increased synthetic data, repeated SFT, and other training choices.

  3. Section 2.4 emphasizes curvature-aware alignment as the main FREE mechanism, yet Table 1 shows that it underperforms simpler gradient- and embedding-based selectors. Moreover, even the best FREE variant (77.6%) falls below the reported SF@3 result (78.3%). Therefore, the additional benefit of FREE over conventional multi-candidate selection remains unclear, particularly given the emphasis on improvements over First-only accuracy.

Questions for authors
  1. How are the seed-relative specifications constructed and verified, and how is structural consistency evaluated in practice? Please clarify the implementation of SpecsynSpec_{syn} VfinalV_{final}, and $S(z',\hat z')$, including the specific validation criteria and procedures.

  2. Can the authors provide more systematic ablation studies to quantify each component's contribution? In particular, comparisons between different validation strategies, random versus influence-based sample selection, and single-stage versus two-stage SFT under comparable training budgets would help clarify the sources of improvement.

  3. Why does the proposed curvature-aware alignment method underperform simpler alternatives, such as gradient alignment and embedding-based selection, in Table 1?

9 comments
AuthorsOpenReview ↗

Response to Reviewer LQGT for Q1

Dear Reviewer LQGT,

Thank you for your thoughtful review. We clarify how the intended specification, structural checks, and retrospective selection are implemented.

1. Constructing the seed-relative specification (Eqs. 8–9).

Γj=Specsyn(rπ(j)0,sj,hj).\Gamma_j=Spec_{syn}(r^0_{\pi(j)},s_j,h_j).

These requirements are supplied through the synthesis prompt: the seed provides a validated problem–code example, the scenario defines the new decision setting, and the scaffold supplies modeling guidance. Equation (9), rj∈Fspec(Γj)r_j\in\mathcal F_{spec}(\Gamma_j), expresses the intended specification-preservation condition.

To support diversity, synthesis may change variables, objectives, and constraints. We therefore impose no universal list of unchanged seed structures. Once generated, z′z' becomes the executable anchor for checking consistency with its textual description and reconstructed code.

2. Evaluating structural consistency S(z′,z^′)S(z',\hat z').

Rule-based checks extract features from solver-API calls and recognizable code patterns. They compare:

  • Objective direction, when recognized in both programs.
  • Binary and integer variable presence, flagging either removal or introduction.
  • Source-dependent structures, including flow or inventory balance, conic structure, subtour elimination, startup, ramping, and sequencing logic.

A required structure detected in z′z' but absent from z^′\hat z' triggers failure. Structures not detected in z′z' impose no additional requirements. Passing SS means that no conflict was detected by these checks; identical code, variable names, and code organization are unnecessary.

3. Applying VfinalV_{final}.

We render z′z' as text x′x' and reconstruct z^′\hat z' through a separate call to the same LLM. Acceptance requires structural consistency, successful execution with finite objective values, and numerical agreement:

\leq \max(10^{-6},10^{-3}\max(|f(z')|,|f(\hat z')|)).$$ When both objective magnitudes are below $10^{-6}$, the absolute difference must instead be at most $10^{-8}$. Structural or execution failures trigger bounded repair and rechecking. A finite objective mismatch normally triggers a new textual rendering from the unchanged anchor. We try up to **five textual views** and require at least one passing reconstruction; otherwise, the candidate is rejected. Accepted experiences form $\mathcal D_V$. On **119 known-good pairs**, multi-view verification retains **49 pairs**, compared with **36** under single-view verification. This recovers **13 additional valid pairs** and reduces false rejection from **69.75% to 58.82%** without relaxing the acceptance criteria. The experiment demonstrates why multiple views help retain more valid synthetic training examples. **4. Connecting seed-relative alignment to retrospective learning.** Trusted problem–code pairs provide the reference objective: $$L_0(\theta)=\frac{1}{|\mathcal D_0|} \sum_{(x,z)\in\mathcal D_0}\ell(f_\theta(x),z).$$ For each verified experience $r_j=(x_j',z_j')$, define its training loss as $$\ell_j(\theta)=\ell(f_\theta(x_j'),z_j').$$ At the broad-training checkpoint $\theta_1$, let $$g_0=\nabla L_0(\theta_1), g_j=\nabla\ell_j(\theta_1).$$ Its approximate influence on trusted-seed loss is $$\widetilde I_0(r_j) =-g_0^\top(G_1+\lambda I)^{-1}g_j,$$ where $G_1$ is the empirical gradient second-moment matrix over the broad-training data and $\lambda>0$ provides regularization. Under this local approximation, increasing the experience’s training weight by a small $\epsilon$ predicts $$\Delta L_0\approx\epsilon\,\widetilde I_0(r_j).$$ We retain the **50% with the smallest estimated influences**, corresponding to the greatest predicted loss reductions. Because the reference targets are validated formulations, this favors experiences expected to reinforce trusted modeling behavior. Curvature accounts for the model’s local response to parameter changes, beyond gradient similarity alone. **5. Scope, scalability, and reproducibility.** Reconstruction checks determine admission to $\mathcal D_V$; retrospective selection then scores each experience against the aggregate trusted-seed loss. Regularized regression avoids explicitly constructing and inverting the full curvature matrix, providing a shared selection criterion across formulation families. We will clarify that Eqs. (8)–(9) define the intended admissible set, while these procedures provide consistency checks and an indirect signal of specification alignment and it is not proof of every requirement in $\Gamma_j$.
AuthorsOpenReview ↗

Response to Reviewer LQGT for Q2

Q2. Ablations of selection and validation

We compare the broad model 1, Random 50%, and Retrospective 50%, with matched subset size and training budget for the latter two. Retrospective selection improves First accuracy on five of six benchmarks, from 66.39% to 69.43% (+3.04 points), without additional inference cost. SF@3 increases from 77.88% to 78.28%;

Macro-average accuracy (%).

Training checkpointFirstSF@3Mean LogLikTrusted Emb.Gradient Align.Curvature Align.Oracle@3
Broad model 162.5868.8269.0867.7168.0266.5270.35
Random 50%66.3977.8877.3378.8078.1077.0581.51
Retrospective top 50% (2)69.4378.2878.5877.7277.4577.6481.42

Unweighted mean across six benchmarks. Each checkpoint generates three candidates per problem: seed 44, temperature 0.1, top-p 0.8. First uses the first candidate; SF@3 uses majority voting; Oracle@3 counts success if any candidate is correct. All selectors rescore the same saved candidates without parameter updates. MAMO uses the cleaned and fixed datasets.

All selectors improve macro-average accuracy over First across the three checkpoints (+3.94–12.41 points), although none consistently outperforms SF@3. Reference-based FREE variants permit bank updates without retraining and provide inspectable reference formulations.

Q2: Detailed results (%)

(a) Broad model 1

DatasetFirstSF@3Mean LogLikTrusted Emb.Gradient Align.Curvature Align.Oracle@3
NL4OPT78.8781.2281.2280.2880.2880.2882.16
NL4LP77.5379.2180.9080.9078.6578.6581.46
MAMO-Easy89.1792.4892.6692.1192.4892.2993.39
MAMO-Complex71.1777.4874.7778.3876.5875.6880.18
IndustryOR30.9538.1040.4835.7135.7133.3340.48
ComplexOR27.7844.4444.4438.8944.4438.8944.44
Macro-average62.5868.8269.0867.7168.0266.5270.35

(b) Random 50%

DatasetFirstSF@3Mean LogLikTrusted Emb.Gradient Align.Curvature Align.Oracle@3
NL4OPT80.2887.3286.3887.3286.8586.3887.32
NL4LP89.8998.3198.3198.3197.1996.6398.31
MAMO-Easy88.6295.4195.6095.9695.0594.8697.25
MAMO-Complex52.2571.1763.0672.9772.0769.3778.38
IndustryOR59.5259.5259.5257.1461.9059.5266.67
ComplexOR27.7855.5661.1161.1155.5655.5661.11
Macro-average66.3977.8877.3378.8078.1077.0581.51

(c) Retrospective top 50% (2)

DatasetFirstSF@3Mean LogLikTrusted Emb.Gradient Align.Curvature Align.Oracle@3
NL4OPT81.6986.8587.7987.7987.3287.7988.26
NL4LP91.0196.6397.1996.6397.1996.0797.19
MAMO-Easy90.6494.8695.6095.4195.4194.5096.88
MAMO-Complex54.0573.8770.2772.9772.0774.7778.38
IndustryOR54.7661.9059.5252.3857.1457.1466.67
ComplexOR44.4455.5661.1161.1155.5655.5661.11
Macro-average69.4378.2878.5877.7277.4577.6481.42

Pooled accuracy (%).

Total correct instances divided by total evaluated instances; each instance receives equal weight.

Training checkpointFirstSF@3Mean LogLikTrusted Emb.Gradient Align.Curvature Align.Oracle@3
Broad model 180.2383.7583.9483.5783.3082.9485.11
Random 50%81.4189.8088.9990.2589.5388.9091.79
Retrospective top 50%83.1289.5389.8089.6289.6289.3591.61
Retrospective − Random (pp)+1.71−0.27+0.81−0.63+0.09+0.45−0.18
AuthorsOpenReview ↗

Response to Reviewer LQGT for Q3

FREE’s central contribution is reusing the verified experience bank for customizable, update-free test-time program selection. FREE-Trusted Embedding, FREE-Gradient Alignment, and FREE-Curvature Alignment provide alternative reference-based scoring rules; FREE-Mean LogLik provides the model-confidence baseline.

1. Role of curvature-aware alignment.

Curvature correction accounts for differences in scale and correlation among reference-gradient directions. It may help when these patterns are informative and reliably estimated. Because the regularized matrix uses at most eight retrieved references, its effectiveness can depend on reference quality and regularization. This is a possible explanation for the observed differences, rather than an experimentally established cause.

In the original Table 1, FREE-Gradient Alignment achieves 77.6%, FREE-Trusted Embedding 77.4%, and FREE-Curvature Alignment 76.19%. The updated Q2 comparison gives a different ordering for the retrospective checkpoint: FREE-Trusted Embedding 77.72%, FREE-Curvature Alignment 77.64%, and FREE-Gradient Alignment 77.45%. We will clarify that curvature-aware alignment is an explored variant, rather than a consistently superior scoring rule.

2. Comparison with self-consistency.

SF@3 achieves 78.3% in the original results and 78.28% in the updated retrospective-checkpoint evaluation, compared with 77.72% for the best reference-based selector in the latter comparison. Dataset-level results show complementary strengths.

Dataset-level win rates against SF@3. Each selector is compared with SF@3 on six benchmarks across three checkpoints. A win means strictly higher accuracy; ties are counted separately. Each dataset–checkpoint comparison receives equal weight.

SelectorBroad model: wins/6Random 50%: wins/6Retrospective 50%: wins/6Overall win rateWins / ties / losses
FREE-Mean LogLik3/6 (50.0%)2/6 (33.3%)4/6 (66.7%)9/18 (50.0%)9 / 4 / 5
FREE-Trusted Embedding2/6 (33.3%)3/6 (50.0%)3/6 (50.0%)8/18 (44.4%)8 / 3 / 7
FREE-Gradient Alignment0/6 (0.0%)2/6 (33.3%)3/6 (50.0%)5/18 (27.8%)5 / 4 / 9
FREE-Curvature Alignment0/6 (0.0%)0/6 (0.0%)2/6 (33.3%)2/18 (11.1%)2 / 3 / 13

For the retrospective checkpoint, FREE-Trusted Embedding and FREE-Gradient Alignment each outperform SF@3 on three benchmarks and tie on one; FREE-Curvature Alignment wins on two and ties on one. At least one reference-based variant wins on five of six benchmarks. This describes coverage across variants, not the performance of a single fixed selector.

Pooled accuracy (%). All selectors rescore the same saved candidates within each checkpoint. Pooled accuracy weights each test instance equally.

CheckpointFirstSF@3FREE-Mean LogLikFREE-Trusted EmbeddingFREE-Gradient AlignmentFREE-Curvature AlignmentOracle@3
Broad model θ1\theta_180.2383.7583.9483.5783.3082.9485.11
Random 50%81.4189.8088.9990.2589.5388.9091.79
Retrospective 50%83.1289.5389.8089.6289.6289.3591.61
Retrospective − Random (pp)+1.71−0.27+0.81−0.63+0.09+0.45−0.18

Pooled accuracy = total correctly solved instances / total evaluated instances. SF@3 denotes self-consistent majority voting over three candidates. FREE-Mean LogLik is the model-confidence baseline; the other three FREE variants use retrieved verified references.

For the retrospective checkpoint, FREE-Trusted Embedding and FREE-Gradient Alignment achieve 89.62%, FREE-Curvature Alignment 89.35%, and SF@3 89.53%. The reference-based selectors are therefore within 0.18 percentage points of SF@3 under this aggregation.

3. Additional benefits of the verified experience bank.

Reference-based FREE complements answer agreement with an external reference signal:

  • Customization: adding domain-relevant or expert-verified experiences can adjust candidate rankings without retraining the model or regenerating candidates.
  • Inspectability: users can examine the retrieved formulations and their modeling structures when reviewing a selection.

A correct numerical answer does not always imply a correct formulation. For example, an omitted constraint may be inactive on one instance but become important when the data change. Reference-based FREE adds a compatibility signal from trusted formulations through embeddings or gradients, encouraging alignment with validated modeling behavior.

Reviewer_JoN4OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Focusing on operations research (OR) modeling, this paper studies how to synthesize structurally diverse and specification-consistent training data from limited trusted seeds for OR modeling. They propose a data augmentation approach with the idea of Retrospective Learning to help OR modeling, RELIVE-OR, which consists of executable-space structural synthesis, closed-loop verification, influence-based experience re-selection, and update-free test-time program selection. The underlying idea is to expand structural coverage while preserving OR model specifications, then revisit verified experiences according to their utility to the current learner. They tested the learned model on six OR modeling benchmarks. The results show that their 8B model achieves 69.4% average single-generation accuracy and 78.3% with three-generation self-consistency, approaching the reported 78.9% of GPT-OSS-120B.

Critical strengths

The paper identifies a meaningful limitation of execution-based verification. A program can execute successfully and return a plausible objective value while violating the intended modeling requirements. The synthesis method targets structural diversity by changing variables, objectives, and constraints in executable programs, rather than only rephrasing questions or changing numerical values.

Critical weaknesses
  1. The necessity of synthetic data is not shown. Downstream performance does not isolate the value of the synthetic data. The reported scores come from a pipeline that also includes data selection, additional fine-tuning, LoRA merging, and test-time program selection. They do not show that the generated data are better than data from other synthesis methods under matched training conditions. Nor do they show that new data are needed: training longer on the original data or reweighting existing samples may achieve similar gains. These controlled comparisons are missing.
  2. The method combines synthesis, verification, influence-based selection, additional fine-tuning, and test-time selection. Controlled ablations are needed to identify which components improve performance.
  3. Synthetic-data quality is only partially evaluated. Structural changes do not establish modeling correctness. The verifier is tested on 119 known-good pairs, without measuring how often it accepts incorrect pairs. It still rejects 58.8% of the known-good pairs.
  4. Generalization beyond the benchmarks is not established. The gains may reflect alignment with the benchmark distributions. Evaluation on unseen problem families and real-world problems is needed to support broader claims.
  5. No code or generated data are provided. The paper also leaves key details unclear, including the 8B base model, the final training-data composition, and the implementation of specification checks. The candidate counts are inconsistent across sections. These gaps prevent reliable reproduction of the reported results from the paper alone.
Questions for authors

See weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer JoN4 for W1

Dear Reviewer JoN4, Thank you for this comment. We clarify the role of synthetic experiences and the evidence for each component.

Response to W1.

1. Why use synthetic data.

The importance of synthetic data for OR modeling has been studied extensively in ORLM, OptMATH, AutoOR, and DeepOR. These methods use synthetic experiences to overcome the scarcity and limited coverage of expert-verified problem–program pairs. ORLM provides direct evidence: macro-average accuracy increases from 51.8% with 686 seed examples to 60.8% with 32,481 training examples.

Our work builds on this established value and asks how to generate and select synthetic experiences that combine diversity, correctness, and specification alignment. All our training conditions include the same trusted seeds, yet their performance differs. The matched subset comparison demonstrates that the choice of synthetic experiences matters beyond seed inclusion alone, supporting retrospective selection as a way to improve their training utility.

2. Comparison with alternative synthesis methods

Table 1 reports accuracy comparisons with ORLM, OptMATH, and AutoOR. Appendix Tables 4–5 compare synthesis, verification, and learning procedures. Our implementation comparison uses Qwen3-8B throughout; we will distinguish implementation results from published reference scores and document training settings.

These construction statistics describe different pipeline stages; they do not themselves establish matched budgets or comparative data quality.

MethodBackbone in our comparisonSeeds / templatesGeneratedVerified / retainedReported training count
ORLMQwen3-8B686 casesRaw count unreported32,48132,481
OptMATHQwen3-8B53 generators>600,000 mathematical instances~150,000 validated + 50,000 augmented~200,000
AutoORQwen3-8B5 LP templates; 9 MILP families; 1 nonlinear template**~29,000–33,000 (estimate)**¹7–8% overall acceptance2,303 summed category/phase entries
RELIVE-ORQwen3-8B119 trusted seeds; distribution in Appendix C.142,958 across rounds27,434 across rounds13,717 selected + 119 seeds = 13,836²

¹ AutoOR's estimate assumes non-overlapping training sets containing all accepted examples; its generation total is not author-reported.

² RELIVE-OR's final count assumes no further filtering/deduplication and seeds added once.

3. What does retrospective selection contribute?

We compare the full verified pool, random 50%, and retrospective 50%; all include the seeds.

Training conditionFirstSF@3Oracle@3
Full verified pool62.5868.8270.35
Random selection (50%)66.3977.8881.51
Retrospective selection (50%)69.4378.2881.42

Unweighted macro-average accuracy (%) across six benchmarks.

Random and retrospective conditions share initialization, subset size, adapter configuration, optimizer settings, and training budget. Retrospective selection improves First by 3.04 percentage points, winning on five of six benchmarks. SF@3 rises by 0.40 points; Oracle@3 is nearly unchanged.

Relative to the full-pool checkpoint, the retrospective condition improves First by 6.85 points, SF@3 by 9.46, and Oracle@3 by 11.07, using half the synthetic experiences. This reflects the overall training procedure; the matched subset comparison more directly assesses selection.

4. Could other components explain the selection gain?

Additional training and adapter configuration are shared by the random and retrospective conditions. First evaluates only the first candidate, without majority voting or FREE, so its improvement does not depend on test-time selection.

We separately compare scoring on the same three saved candidates within each checkpoint, without parameter updates. Across three checkpoints, all four scoring variants improve macro-average accuracy over First by 3.94–12.41 points, approaching the Oracle@3 candidate-set upper bound.

5. What about longer seed-only training or reweighting?

Seed inclusion alone cannot explain the matched selection gain. A seed-only or reweighting control would separately quantify the incremental benefit of adding synthetic examples. We will distinguish complete-method comparisons from the selection ablation, rather than attribute the full pipeline gain solely to synthesis.

References

  • ORLM: A Customizable Framework in Training Large Models for Automated Optimization Modeling
  • OptMATH: A Scalable Bidirectional Data Synthesis Framework for Optimization Modeling
  • AutoOR: Scalably Post-training LLMs to Autoformulate Operations Research Problems
  • DeepOR: A Deep Reasoning Foundation Model for Optimization Modeling
AuthorsOpenReview ↗

Response to Reviewer JoN4 for W2.2

Response to W2.2 (Controlled ablations are needed to identify which components improve performance.)

We compare the broad model θ₁ (full verified pool), Random 50%, and Retrospective 50%, with matched subset size and training budget. All three Qwen3-8B checkpoints include the trusted seeds.

1. Training-data selection. Retrospective selection improves First accuracy on five of six benchmarks, raising the macro-average from 66.39% to 69.43% (+3.04 percentage points) without additional inference cost. SF@3 increases from 77.88% to 78.28%.

Both 50% conditions also outperform the broad checkpoint in aggregate: SF@3 increases from 68.82% to 77.88–78.28%, and Oracle@3 from 70.35% to 81.42–81.51%. This comparison reflects the overall training procedure; the matched subset comparison more directly assesses retrospective selection.

2. Test-time selection. All four selectors rescore the same three saved candidates within each checkpoint, without parameter updates. Each improves macro-average accuracy over First across all three checkpoints, by 3.94–12.41 points. Relative performance varies; no selector consistently dominates SF@3. The reference-based variants additionally support an updatable bank and inspectable retrieved formulations.

Structural-diversity and retention analyses separately assess synthesis and verification; the comparisons below assess training-data selection and test-time scoring.

Macro-average accuracy (%).

ModelFirstSF@3Mean LogLikEmbeddingGradientCurvatureOracle@3
Full pool (θ₁)62.5868.8269.0867.7168.0266.5270.35
Random 50%66.3977.8877.3378.8078.1077.0581.51
Retrospective 50%69.4378.2878.5877.7277.4577.6481.42

Unweighted mean across six benchmarks. First: first candidate; SF@3: majority voting; Oracle@3: any correct candidate. Mean LogLik, Embedding, Gradient, and Curvature denote FREE-Mean LogLik, FREE-Trusted Embedding, FREE-Gradient Alignment, and FREE-Curvature Alignment. Only the latter three use reference memory.

Pooled accuracy (%).

ModelFirstSF@3Mean LogLikEmbeddingGradientCurvatureOracle@3
Full pool (θ₁)80.2383.7583.9483.5783.3082.9485.11
Random 50%81.4189.8088.9990.2589.5388.9091.79
Retrospective 50%83.1289.5389.8089.6289.6289.3591.61
Retrospective − Random+1.71−0.27+0.81−0.63+0.09+0.45−0.18

Differences are percentage points.

Per-dataset accuracy (%).

(a) Full verified pool (θ₁)

DatasetFirstSF@3Mean LogLikEmbeddingGradientCurvatureOracle@3
NL4Opt78.8781.2281.2280.2880.2880.2882.16
NL4LP77.5379.2180.9080.9078.6578.6581.46
MAMO Easy89.1792.4892.6692.1192.4892.2993.39
MAMO Complex71.1777.4874.7778.3876.5875.6880.18
IndustryOR30.9538.1040.4835.7135.7133.3340.48
ComplexOR27.7844.4444.4438.8944.4438.8944.44
Macro-average62.5868.8269.0867.7168.0266.5270.35

(b) Random selection (50%)

DatasetFirstSF@3Mean LogLikEmbeddingGradientCurvatureOracle@3
NL4Opt80.2887.3286.3887.3286.8586.3887.32
NL4LP89.8998.3198.3198.3197.1996.6398.31
MAMO Easy88.6295.4195.6095.9695.0594.8697.25
MAMO Complex52.2571.1763.0672.9772.0769.3778.38
IndustryOR59.5259.5259.5257.1461.9059.5266.67
ComplexOR27.7855.5661.1161.1155.5655.5661.11
Macro-average66.3977.8877.3378.8078.1077.0581.51

(c) Retrospective selection (50%)

DatasetFirstSF@3Mean LogLikEmbeddingGradientCurvatureOracle@3
NL4Opt81.6986.8587.7987.7987.3287.7988.26
NL4LP91.0196.6397.1996.6397.1996.0797.19
MAMO Easy90.6494.8695.6095.4195.4194.5096.88
MAMO Complex54.0573.8770.2772.9772.0774.7778.38
IndustryOR54.7661.9059.5252.3857.1457.1466.67
ComplexOR44.4455.5661.1161.1155.5655.5661.11
Macro-average69.4378.2878.5877.7277.4577.6481.42
AuthorsOpenReview ↗

Response to Reviewer JoN4 for W2.1

Response to W2.1 (The method combines synthesis, verification, influence-based selection, additional fine-tuning, and test-time selection.)

2.1 Contribution of each component.

RELIVE-OR addresses one problem: synthesis needs structural freedom to go beyond the structures covered by the trusted seeds, but more freedom raises the risk of violating their OR model specification (variable roles, constraint logic, objective semantics). Each component addresses one part of this problem, and we report controlled evidence for each.

(1) Synthesis in executable code space (DE-SCALE). Prior OR synthesis pipelines mainly diversify the natural-language problem or the mathematical formulation (ORLM, ReSocratic), or vary data within fixed formulation templates (OptMATH, AutoOR). MILP-Evolve evolves executable MILP code, but to train GNN solvers rather than to build language–program pairs. To our knowledge, DE-SCALE is the first method to evolve the executable program itself to create natural-language–program pairs for LLM autoformulation. We work in code space because code states modeling structure explicitly and can be checked by execution, whereas natural language often leaves it implicit.

Evidence of structural diversity (Appendix D, 6,539 randomly sampled seed–variant pairs):

  • 100% of variants have a new executable AST.
  • 95.47% are structural formulation variants, and a further 4.53% shift to a different OR formulation family.
  • The mean normalized structural-AST similarity to the seed is 0.585.
  • 0% are identical code, surface-only rewrites, or instance-only perturbations.

(2) Multi-view verification and repair. Single-view back-translation rejects a valid program whenever its one description is lossy. We keep the synthesized program fixed, generate up to five independent descriptions, and accept the program if any one reconstructs it. On 119 known-good pairs, this recovers 13 additional valid pairs and lowers false rejection from 69.75% to 58.82%, without any code repair (Appendix E).

(3) Retrospective influence selection. Influence-based data selection is not new, and we do not claim it is. What differs is when it is applied: we score experiences that are already verified, after broad training on the full pool, against a trusted reference set, and then train on the selected subset. In a controlled comparison where the random and influence-selected subsets share the same initialization, subset size, adapters, and optimizer steps, influence selection improves first-sample accuracy from 66.39% to 69.43% (macro-average), on 5 of 6 benchmarks. Both selected subsets also outperform the full pool (62.58%) despite less training.

(4) Update-free test-time selection (FREE). To our knowledge, this is the first use of a verified synthetic experience bank as a non-parametric reference memory for choosing among executable candidates in OR autoformulation, without parameter updates. On the same three candidates per problem, reference-based selection raises accuracy over the first sample from 69.43% to 77.45–77.72%, comparable to majority voting (78.28%). FREE does not exceed majority voting in accuracy; its distinctive property is that selection uses verified external references, which can be extended without retraining the model.

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

RELIVE-OR proposes an experience-centric framework for LLM-based operations-research modeling. It treats each problem-program pair as a modeling experience and organizes around acquiring, validating, revisiting, and reusing such experiences. It combines DE-SCALE for generating and validating structually diverse optimization problems, retrospective influence-based selection for focused training, and FREE for retrieving verified experience to guide test-time candidate selection. Experiments across 6 benchmarks demonstrated competitive performance with an 8B model.

Critical strengths

S1. Coherent experience centric framework. RELIVE-OR integrates synthetic data generation, verification, restropective learning, and inference-time reuse into a unified framework. Distinguishing experience validity from learner-dependent usefulness is conceptually meaningful.

S2. Strong empirical anylysis of structural diversity and candidate selection. The analysis of 6539 synthesized examples shows that 95.47% are structural formulation variants, supporting the claim that DE-SCALE goes beyond simple paraphrasing. The gap between First-only(69.4%) and Oracle@3(81.4%) also demonstrates the importance of candiadte selection.

Critical weaknesses

W1. Insufficient validation of seed-relative specification preservation Although this paper formally defines specification preservation and describes several verifiable procedures, the operational implementation of the seed-relative specification check remains unclear. The structural-diversity results demonstrate variation but do not establish semantic preservation. Additional implementation details and direct validation would strengthen this central claim.

W2. Retrospective selection is not isolated from additional training. The focused model differs from the broad model in both data selection and training duration. Without a controlled comparison against an equally sized, non-retrospectively selcted subset under the same training budget, it is unclear whether improvements arise influence-based selection or simply additonal optimizaiton.

Questions for authors

Q1. How is seed-relative specification preservation verified in practice? Which properties must remain invariant, and how are violations detected?

Q2. Can the authors compare the broad model, the restropectively selected focused model, and a same budget focused model trained on a non-retrospectively selected subset?

Q3. Curvature Alignment(76.19%)underperforms gradient alignment (77.6%) in macro-average accuracy. Under what conditions is the curvature correction expected to help?

3 comments
AuthorsOpenReview ↗

Response to Reviewer bRqB for Q1

Dear Reviewer bRqB,

Thank you for your time and thoughtful review. We address your concerns below.

Response to Q1. We clarify the practical implementation of seed-relative specification alignment: consistency checks determine acceptance during synthesis, and curvature-aware retrospective selection subsequently assesses alignment with trusted seeds.

1. How is the intended specification constructed?

The seed, scenario, and scaffold jointly specify the intended requirements through the synthesis prompt:

Γj=Specsyn(rπ(j)0,sj,hj).\Gamma_j=Spec_{syn}(r^0_{\pi(j)},s_j,h_j).

The seed supplies a validated problem–code example, the scenario defines the new decision setting, and the scaffold provides modeling guidance. Equations (8)–(9) describe the intended condition that the synthesized experience belongs to Fspec(Γj)\mathcal F_{spec}(\Gamma_j).

2. Which properties are invariant?

We allow scenario-driven changes to variables, objectives, and constraints, so no universal set of structures must remain identical between the original seed and the synthesized program.

Once z′z' is generated, however, it becomes the executable anchor. Its textual description and reconstructed program z^′\hat z' should preserve its objective direction, variable-domain requirements, and relevant constraint relationships. Names and code organization may differ.

3. How are violations detected?

Acceptance requires three checks:

  • Structural consistency: rule-based solver-API and code-pattern checks flag conflicting objective directions when both are recognized, changes in binary/integer-variable presence, and missing source-dependent structures such as flow balance, conic constraints, and subtour elimination. A structure detected in z′z' must also be detected in z^′\hat z'; structures not detected in z′z' impose no additional requirements.
  • Successful execution: both programs execute and return finite objective values.
  • Numerical agreement: their objectives satisfy
\leq \max(10^{-6},10^{-3}\max(|f(z')|,|f(\hat z')|)).$$ When both objective magnitudes are below $10^{-6}$, the absolute difference must instead be at most $10^{-8}$. Structural or execution failures trigger bounded repair and rechecking. A finite objective mismatch normally triggers another textual rendering from the unchanged anchor. We try up to **five views** and require at least one passing reconstruction; otherwise, the candidate is rejected. These checks assess detectable structural and numerical consistency, without requiring identical code or establishing equality of the complete feasible sets. **4. What does retrospective influence contribute?** Our central contribution is to connect experience selection to **the modeling behavior that training is expected to reinforce**. Reconstruction checks assess consistency within each synthesized experience; retrospective influence additionally evaluates its predicted training effect against trusted problem–code pairs. After broad SFT, we estimate how increasing each verified experience’s training weight would affect trusted-seed loss. The estimate combines the experience gradient and trusted-seed gradient through a **regularized curvature approximation**, accounting for the model’s local response to parameter changes. We retain the **50% predicted to reduce trusted-seed loss the most**. This makes trusted formulations an active selection criterion: experiences are selected for their predicted contribution to learning validated modeling behavior, while allowing diverse scenarios and implementations. The resulting signal goes beyond textual similarity and exact structural matching. **5. Why is this feasible at scale?** Retrospective influence provides a shared selection criterion across OR formulation families, without requiring a separate hand-written specification checker for each family at this stage. Regularized regression avoids explicitly constructing and inverting the full curvature matrix. The contribution is therefore the integration of **reconstruction-based consistency checking with curvature-aware selection guided by trusted formulations**. Reconstruction determines entry into $\mathcal D_V$; retrospective influence then selects experiences predicted to reinforce trusted modeling behavior. This supports specification alignment through training utility, without treating influence as a direct test of every requirement in $\Gamma_j$.
AuthorsOpenReview ↗

Response to Reviewer bRqB for Q2

Q2. Ablations of selection and validation

We compare the broad model 1, Random 50%, and Retrospective 50%, with matched subset size and training budget for the latter two. Retrospective selection improves First accuracy on five of six benchmarks, from 66.39% to 69.43% (+3.04 points), without additional inference cost. SF@3 increases from 77.88% to 78.28%;

Macro-average accuracy (%).

Training checkpointFirstSF@3Mean LogLikTrusted Emb.Gradient Align.Curvature Align.Oracle@3
Broad model 162.5868.8269.0867.7168.0266.5270.35
Random 50%66.3977.8877.3378.8078.1077.0581.51
Retrospective top 50% (2)69.4378.2878.5877.7277.4577.6481.42

Unweighted mean across six benchmarks. Each checkpoint generates three candidates per problem: seed 44, temperature 0.1, top-p 0.8. First uses the first candidate; SF@3 uses majority voting; Oracle@3 counts success if any candidate is correct. All selectors rescore the same saved candidates without parameter updates. MAMO uses the cleaned and fixed datasets.

All selectors improve macro-average accuracy over First across the three checkpoints (+3.94–12.41 points), although none consistently outperforms SF@3. Reference-based FREE variants permit bank updates without retraining and provide inspectable reference formulations.

Q2: Detailed results (%)

(a) Broad model 1

DatasetFirstSF@3Mean LogLikTrusted Emb.Gradient Align.Curvature Align.Oracle@3
NL4OPT78.8781.2281.2280.2880.2880.2882.16
NL4LP77.5379.2180.9080.9078.6578.6581.46
MAMO-Easy89.1792.4892.6692.1192.4892.2993.39
MAMO-Complex71.1777.4874.7778.3876.5875.6880.18
IndustryOR30.9538.1040.4835.7135.7133.3340.48
ComplexOR27.7844.4444.4438.8944.4438.8944.44
Macro-average62.5868.8269.0867.7168.0266.5270.35

(b) Random 50%

DatasetFirstSF@3Mean LogLikTrusted Emb.Gradient Align.Curvature Align.Oracle@3
NL4OPT80.2887.3286.3887.3286.8586.3887.32
NL4LP89.8998.3198.3198.3197.1996.6398.31
MAMO-Easy88.6295.4195.6095.9695.0594.8697.25
MAMO-Complex52.2571.1763.0672.9772.0769.3778.38
IndustryOR59.5259.5259.5257.1461.9059.5266.67
ComplexOR27.7855.5661.1161.1155.5655.5661.11
Macro-average66.3977.8877.3378.8078.1077.0581.51

(c) Retrospective top 50% (2)

DatasetFirstSF@3Mean LogLikTrusted Emb.Gradient Align.Curvature Align.Oracle@3
NL4OPT81.6986.8587.7987.7987.3287.7988.26
NL4LP91.0196.6397.1996.6397.1996.0797.19
MAMO-Easy90.6494.8695.6095.4195.4194.5096.88
MAMO-Complex54.0573.8770.2772.9772.0774.7778.38
IndustryOR54.7661.9059.5252.3857.1457.1466.67
ComplexOR44.4455.5661.1161.1155.5655.5661.11
Macro-average69.4378.2878.5877.7277.4577.6481.42

Pooled accuracy (%).

Total correct instances divided by total evaluated instances; each instance receives equal weight.

Training checkpointFirstSF@3Mean LogLikTrusted Emb.Gradient Align.Curvature Align.Oracle@3
Broad model 180.2383.7583.9483.5783.3082.9485.11
Random 50%81.4189.8088.9990.2589.5388.9091.79
Retrospective top 50%83.1289.5389.8089.6289.6289.3591.61
Retrospective − Random (pp)+1.71−0.27+0.81−0.63+0.09+0.45−0.18
AuthorsOpenReview ↗

Response to Reviewer bRqB for Q3

Thank you for highlighting this distinction. FREE’s central contribution is reusing the verified experience bank for customizable, update-free test-time program selection. FREE-Trusted Embedding, FREE-Gradient Alignment, and FREE-Curvature Alignment provide alternative reference-based scoring rules; FREE-Mean LogLik provides the model-confidence baseline.

1. Role of curvature-aware alignment.

Curvature correction accounts for differences in scale and correlation among reference-gradient directions. It may help when these patterns are informative and reliably estimated. Because the regularized matrix uses at most eight retrieved references, its effectiveness can depend on reference quality and regularization. This is a possible explanation for the observed differences, rather than an experimentally established cause.

In the original Table 1, FREE-Gradient Alignment achieves 77.6%, FREE-Trusted Embedding 77.4%, and FREE-Curvature Alignment 76.19%. The updated Q2 comparison gives a different ordering for the retrospective checkpoint: FREE-Trusted Embedding 77.72%, FREE-Curvature Alignment 77.64%, and FREE-Gradient Alignment 77.45%. We will clarify that curvature-aware alignment is an explored variant, rather than a consistently superior scoring rule.

2. Comparison with self-consistency.

SF@3 achieves 78.3% in the original results and 78.28% in the updated retrospective-checkpoint evaluation, compared with 77.72% for the best reference-based selector in the latter comparison. Dataset-level results show complementary strengths.

Dataset-level win rates against SF@3. Each selector is compared with SF@3 on six benchmarks across three checkpoints. A win means strictly higher accuracy; ties are counted separately. Each dataset–checkpoint comparison receives equal weight.

SelectorBroad model: wins/6Random 50%: wins/6Retrospective 50%: wins/6Overall win rateWins / ties / losses
FREE-Mean LogLik3/6 (50.0%)2/6 (33.3%)4/6 (66.7%)9/18 (50.0%)9 / 4 / 5
FREE-Trusted Embedding2/6 (33.3%)3/6 (50.0%)3/6 (50.0%)8/18 (44.4%)8 / 3 / 7
FREE-Gradient Alignment0/6 (0.0%)2/6 (33.3%)3/6 (50.0%)5/18 (27.8%)5 / 4 / 9
FREE-Curvature Alignment0/6 (0.0%)0/6 (0.0%)2/6 (33.3%)2/18 (11.1%)2 / 3 / 13

For the retrospective checkpoint, FREE-Trusted Embedding and FREE-Gradient Alignment each outperform SF@3 on three benchmarks and tie on one; FREE-Curvature Alignment wins on two and ties on one. At least one reference-based variant wins on five of six benchmarks. This describes coverage across variants, not the performance of a single fixed selector.

Pooled accuracy (%). All selectors rescore the same saved candidates within each checkpoint. Pooled accuracy weights each test instance equally.

CheckpointFirstSF@3FREE-Mean LogLikFREE-Trusted EmbeddingFREE-Gradient AlignmentFREE-Curvature AlignmentOracle@3
Broad model θ1\theta_180.2383.7583.9483.5783.3082.9485.11
Random 50%81.4189.8088.9990.2589.5388.9091.79
Retrospective 50% (θ2\theta_2)83.1289.5389.8089.6289.6289.3591.61
Retrospective − Random (pp)+1.71−0.27+0.81−0.63+0.09+0.45−0.18

Pooled accuracy = total correctly solved instances / total evaluated instances. SF@3 denotes self-consistent majority voting over three candidates. FREE-Mean LogLik is the model-confidence baseline; the other three FREE variants use retrieved verified references.

For the retrospective checkpoint, FREE-Trusted Embedding and FREE-Gradient Alignment achieve 89.62%, FREE-Curvature Alignment 89.35%, and SF@3 89.53%. The reference-based selectors are therefore within 0.18 percentage points of SF@3 under this aggregation. These results show that the comparison depends on the dataset, checkpoint, and aggregation metric.

3. Additional benefits of the verified experience bank.

Reference-based FREE complements answer agreement with an external reference signal:

  • Customization: adding domain-relevant or expert-verified experiences can adjust candidate rankings without retraining the model or regenerating candidates.
  • Inspectability: users can examine the retrieved formulations and their modeling structures when reviewing a selection.

A correct numerical answer does not always imply a correct formulation. For example, an omitted constraint may be inactive on one instance but become important when the data change. Reference-based FREE adds a compatibility signal from trusted formulations through embeddings or gradients, encouraging alignment with validated modeling behavior.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal Phase

Dear Area Chair, Program Chairs, and Reviewers,

We sincerely thank you for your time and constructive feedback. Below, we summarize the main concerns raised during the review, our responses, and the corresponding manuscript revisions. We welcome any further comments or questions.

Reviewers’ concernsResponses and revisions
Connection to the central objective (LQGT, JoN4)Section 2 now connects the components to specification-faithful OR modeling: DE-SCALE and validation screen incorrect supervision; retrospective learning prioritizes experiences predicted to improve trusted-formulation learning; FREE guides final candidate selection. Appendix A.2 provides the conditional connection between distributional mismatch, reference-loss improvement, and specification error.
Specification checks and verifier reliability (bRqB, LQGT, JoN4)Section 2.2 specifies structural checks, execution criteria, numerical tolerances, multi-view reconstruction, and bounded repair. The validation ablation reduces false rejection from 69.75% to 58.82% on 119 known-good pairs. This measures retention of valid supervision.
Selection gains versus additional training (all three reviewers)A matched-budget comparison of Random 50% and Retrospective 50% isolates the selection effect. Retrospective selection improves first-candidate accuracy from 66.39% to 69.43%, winning on five of six benchmarks. Section 4 and Appendix D.4 report the results.
Evidence beyond objective agreement (LQGT, JoN4)Section 4 and Appendix D.3 add two-judge formulation assessments, gated and ungated scores, and paired examples where both programs return the correct objective but differ in constraint fidelity.
FREE and curvature-aware scoring (bRqB, LQGT)Sections 2.4 and B.3 clarify FREE’s contribution: reference-based compatibility scoring of generated programs without parameter updates. Fixed-candidate comparisons report embedding, gradient, curvature, likelihood, and majority-vote results; curvature is presented as an alternative, not a consistently superior selector.
Reproducibility and generalization (JoN4)Appendix C consolidates the backbone, data roles, training and adapter settings, reference-bank construction, and three-candidate evaluation protocol. The manuscript specifies the benchmark evaluation scope and code/data release plan.

These revisions distinguish controlled selection evidence from full-pipeline comparisons and connect the implementation, theory, and formulation-level evaluation to the central objective. Reviewer LQGT stated that most concerns had been addressed and increased the rating to 3. The revisions also respond to overlapping concerns raised by Reviewer JoN4 on component contributions and evaluation beyond objective agreement, and Reviewer bRqB on specification checks, selection ablations, and FREE scoring. We have not yet received follow-up assessments and welcome further feedback.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.