JOR-Bench: Japanese Operations Research Benchmarks for Large Language Models

Yuu Jinnai

REVIEW 3 major objections 6 minor 11 references

Large language models formulate and solve operations-research problems about equally well in English and Japanese, with only a 0.3-point accuracy gap, but they fail in different ways depending on the language.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 19:57 UTC pith:JHHPU2V4

load-bearing objection A useful new Japanese OR evaluation resource with a plausible language-neutrality result; the main risk is unverified translation fidelity, but the error analysis is well-executed and the paper deserves a serious referee. the 3 major comments →

arxiv 2607.16777 v1 pith:JHHPU2V4 submitted 2026-07-18 cs.CL cs.AI

JOR-Bench: Japanese Operations Research Benchmarks for Large Language Models

Yuu Jinnai This is my paper

classification cs.CL cs.AI

keywords operations researchlarge language modelsJapanese benchmarkcross-lingual evaluationoptimization modelingLLM code generationbenchmark translationOR-Tools

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

JOR-Bench translates five existing English operations-research (OR) benchmarks — 1,319 problems in linear, mixed-integer, nonlinear, and combinatorial optimization — into Japanese, pairing each Japanese statement with its expected numerical answer. Evaluating seven LLMs on both language versions, the paper finds that OR formulation ability is largely language-neutral: the average Japanese-minus-English accuracy difference is only −0.3 percentage points, and 28 of 35 model–dataset pairs are statistically indistinguishable from zero. The aggregate hides a real story, though. Japanese phrasing can push models to output decision-variable values instead of the objective value, and some models follow the instruction to emit a Python code block less reliably in Japanese. The paper's point is that English performance alone does not predict Japanese performance, so non-English practitioners need benchmarks in their own language.

Core claim

On the paper's own terms, the central claim is that OR formulation ability, measured as code-generation accuracy, is largely language-neutral: across seven models and five benchmarks (1,319 problems), the average Japanese-minus-English accuracy is only −0.3 pp, and 28 of 35 model–dataset pairs are statistically indistinguishable from zero. The discovery is that the aggregate hides systematic cross-lingual differences: in Japanese, models on NL4OPT and OptiBench often output decision-variable values instead of the objective value (a pragmatic disambiguation failure), and on MAMO they generate Python code blocks less reliably. Per-problem agreement analysis shows that on the hardest combinator

What carries the argument

The load-bearing mechanism is JOR-Bench itself: 1,319 Japanese problem statements paired with expected numerical answers, created by LLM-based translation followed by human post-editing of five English OR benchmarks. The evaluation protocol makes language the only variable: every model receives the same prompt template, generates a standalone Python script, and is scored by executing that script with OR-Tools in a sandbox (300-second timeout, 32GB memory) and comparing the numerical output to the expected answer within 10^-6. This execution-based design, plus paired statistical tests (McNemar's test, Cohen's κ), isolates genuine language effects from noise and reveals that per-problem behavi

Load-bearing premise

The Japanese test set is valid: that a single post-editor's LLM-assisted translation of 1,319 problems preserves the mathematical semantics, constraints, tables, and expected answers well enough that the measured accuracy differences reflect model behavior rather than translation artifacts.

What would settle it

Re-translate a random sample (say 200 problems) of JOR-Bench with an independent human translator, or back-translate the Japanese statements to English and check that the expected numerical answers still match; if the grand average Japanese-minus-English accuracy for the same seven models moves by more than a few percentage points, the language-neutrality result is an artifact of the original translation choices.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Practitioners working in Japanese can expect roughly the same OR formulation accuracy from strong multilingual models as from English prompts, making Japanese-language deployment viable without a major accuracy penalty.
English-only evaluation is insufficient for predicting Japanese behavior: on hard combinatorial problems, per-problem agreement can be near zero, so a model that fails a problem in English may solve it in Japanese and vice versa.
The pragmatic disambiguation failure — outputting decision-variable counts instead of the objective value — suggests that Japanese prompts for optimization problems may need to explicitly state which quantity to report, a cheap prompt-engineering fix.
The no-code failure mode implies that instruction-following degrades in non-primary languages, so code-generation benchmarks should track whether a model produces any code block, not just whether the code is correct.
JOR-Bench provides a reusable, solver-independent resource for future cross-lingual OR research, since any solver or programming language can be evaluated against the same Japanese problem–answer pairs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The finding that English and Japanese unlock complementary problem subsets on hard combinatorial problems suggests a testable extension: prompting the same problem in both languages and taking the union of solutions could raise overall accuracy for models at the edge of capability.
The language-neutrality result may partly reflect that the Japanese benchmark is a translation of English discourse rather than naturally occurring Japanese; a benchmark written originally in Japanese business or engineering language might show larger language effects, as the paper itself concedes.
The single post-editor translation pipeline is the fragile point; an independent replication with a second translator, or a back-translation check, would tell whether the −0.3 pp gap is a property of models or of the specific Japanese wording chosen.
The same cross-lingual evaluation design could be applied to other languages — Chinese, Spanish, Arabic — to see whether the pragmatic disambiguation failure is specific to Japanese or a general phenomenon in non-English prompts.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

A useful new Japanese OR evaluation resource with a plausible language-neutrality result; the main risk is unverified translation fidelity, but the error analysis is well-executed and the paper deserves a serious referee.

read the letter

Worth a read if you care about multilingual evaluation of LLMs for optimization. The paper translates five existing English OR benchmarks into Japanese (1,319 problems), runs seven models on both languages, and finds average accuracy nearly identical (−0.3 pp). The new resource itself is the main contribution — there is no existing Japanese OR benchmark of this scope. The execution protocol is concrete and fair: code must run in a sandbox with OR-Tools, exact output matching, timeouts. The statistical work is appropriate: McNemar per cell, Cohen's κ for per-problem agreement. The error analysis is genuinely useful — the pragmatic disambiguation failure (models output decision-variable values instead of the objective in Japanese) is a real, well-illustrated phenomenon, and the κ analysis showing complementarity on OptMATH is a nice catch.

Soft spots, in order of severity. First, translation validity is under-verified. §2.2 says gpt-oss-120b translated, first author post-edited. No inter-annotator check, no back-translation, no released quality metrics. That is a real gap for a benchmark whose central claim is cross-lingual neutrality. That said, the error analysis partially mitigates the concern: the main failures are about output formatting, not misformulated constraints, and the example translation is faithful. Still, an independent verification step (or at least a back-translation report) would materially strengthen the paper. Second, mild circularity: the translation model is also an evaluated model (GPT-OSS). Not damning — all models saw the same Japanese prompts — but worth acknowledging and ideally controlling for. Third, accuracy numbers are single runs without confidence intervals; McNemar handles significance, but point estimates could use bootstrap intervals. Minor: no URL or commit hash for the released code, and one citation placeholder.

The paper is not trying to break new theoretical ground; it is an evaluation resource with careful empirical analysis. For anyone working on multilingual LLM evaluation or OR automation for Japanese, this is directly useful. The central claim holds up to the extent the translation is trustworthy, which is the one thing I would want verified before relying on the −0.3 pp number.

Recommendation: send it to peer review. It deserves referee time; expect a request for translation verification and minor reproducibility fixes.

Referee Report

3 major / 6 minor

Summary. JOR-Bench introduces a Japanese-language set of 1,319 operations-research problems obtained by translating five established English benchmarks: IndustryOR, MAMO Complex LP, NL4OPT, OptiBench, and OptMATH. Seven LLMs, including Japanese-specialized and multilingual models, are evaluated on both the English and Japanese versions using an execution-based protocol in which each model generates a standalone Python script that is run with OR-Tools and compared to the original numerical answer. The paper's central claim is that OR formulation ability is largely language-neutral for strong multilingual models, with a grand average accuracy difference of only −0.3 pp between English and Japanese; per-model/dataset differences are tested with McNemar's test, and per-problem English/Japanese agreement is quantified with Cohen's kappa. An appendix reports MAMO robustness under PuLP and Pyomo.

Significance. The benchmark fills a real gap, since most OR formulation benchmarks are English-only, and the execution-based protocol with OR-Tools, the use of paired statistical tests, and the release of data and code are concrete strengths. If the Japanese translations are semantically faithful, the near-zero aggregate difference is a useful and non-obvious empirical result, and the per-problem churn on hard combinatorial problems revealed by Cohen's kappa is interesting. However, the central language-neutrality claim is only as strong as the translation fidelity, and the manuscript currently provides no verification that the 1,319 Japanese prompts preserve the mathematical semantics, answer pragmatics, and expected numerical answers of the sources.

major comments (3)

[§2.2, §4.2] Translation fidelity is load-bearing and unverified. The entire Japanese-language evaluation and the −0.3 pp language-neutrality claim depend on the assumption that the Japanese prompts are semantically equivalent to the English originals. The manuscript reports only gpt-oss-120b translation followed by first-author post-editing, with no inter-annotator reliability, no back-translation check, and no released translation-quality metrics. The problem is not merely stylistic: the illustrative example in §4.2 has a Japanese question that asks "ボートとカヌーをそれぞれ何回使うべきでしょうか" (how many times should each be used?) while the expected answer is 1160.0, the total time. By the wording of the Japanese prompt, the model's output (12 and 23) is a plausible answer to the question actually asked, suggesting the "pragmatic disambiguation failure" may be a translation artifact. Please add a verification step: i
[§4, Table 5, Abstract] The aggregate claim "statistically indistinguishable from zero" is not directly supported. McNemar tests are applied to individual model–dataset pairs, but the overall −0.3 pp grand average is reported without a confidence interval or an aggregate hypothesis test. In addition, all accuracy numbers come from a single generation per model–problem, and no decoding parameters (temperature, top-p, max tokens) or number of runs are reported; if sampling is stochastic, single-run accuracies have nontrivial variance. Please report Wilson confidence intervals for the accuracy cells and for the aggregate ΔJA, specify the decoding protocol or run multiple seeds, and provide a paired bootstrap or equivalent test for the overall near-zero difference.
[§2.2, Table 2] There is an overlap between the translation model and the evaluated model suite: gpt-oss-120b is used to create the Japanese prompts and is also evaluated as GPT-OSS. This creates a contamination risk, since the Japanese prompts may be inadvertently aligned with that model's preferred phrasing. The observed GPT-OSS ΔJA (+1.2 pp) does not show obvious inflation, but the design should be decontaminated or tested: re-translate a random subset of problems with a model that is not in the evaluation set, or with human translation, and confirm that the main language-neutrality result and the MAMO/NL4OPT error patterns are unchanged.

minor comments (6)

[Abstract vs. Table 5] The abstract reports an overall difference of −0.3 pp, but the final "Avg" row of Table 5 shows −0.2 pp. Reconcile the rounding or state the exact value consistently.
[Table 5 caption] The abbreviations NL4, OptiB, and OptM are not defined in the caption; define them as NL4OPT, OptiBench, and OptMATH.
[§4.3] The claim that "78% of failures in both languages produce decision-variable output" for GPT-OSS on NL4OPT is not backed by a counts table. Add a small table or appendix giving the failure counts and percentages.
[§3.1 / Reproducibility] The reproducibility statement reports library versions but not the decoding settings. State whether greedy decoding was used, and if not, report temperature, top-p, max tokens, and the number of generations per problem.
[References] The OptMATH reference contains a typo: "InF orty-second" should be "In Forty-second."
[§6] The limitations section acknowledges that the benchmark is "not culturally or contextually Japanese" but does not address semantic fidelity of the translation. Add an explicit limitation or reference to the translation verification analysis.

Circularity Check

0 steps flagged

No significant circularity: the benchmark evaluation is measured against fixed external answers; the translation-model overlap is a validity confound, not a circular reduction.

full rationale

The paper's central claim—that OR formulation ability is largely language-neutral (mean ΔJA = −0.3 pp)—is an empirical measurement, not a derivation from its construction. JOR-Bench inherits all expected numerical answers from five existing English benchmarks (IndustryOR, MAMO Complex LP, NL4OPT, OptiBench, OptMATH), and accuracy is scored by executing model-generated Python code against those fixed answers. No parameter is fitted to the Japanese outcomes, and the Japanese accuracy numbers are not used to construct the Japanese prompts or answers; hence there is no prediction-equals-fit or self-definitional circularity. The Japanese test set is produced by gpt-oss-120b and then post-edited by the first author (§2.2), and GPT-OSS is also one of the seven evaluated models (Table 2). This overlap is a genuine validity concern—the Japanese prompt distribution may be idiosyncratic to the translating model, and no inter-annotator or back-translation check is reported—but it is not a circular reduction: the language-neutrality claim is not defined in terms of GPT-OSS's outputs, and the same translation is used for all models, so the aggregate comparison does not reduce to a fitted input. The only self-citation (CAT-Thinking, Jinnai 2026) identifies the provenance of an evaluated model and is not load-bearing for the paper's conclusions. Section 6 honestly acknowledges that JOR-Bench is 'Japanese-language but not culturally or contextually Japanese,' further indicating awareness of translation limits. These issues are correctness/validity risks, not circularity, and should be addressed by future verification work rather than by flagging a circular derivation.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

No free constants are fitted; the ledger entries are evaluation-protocol choices and assumptions about benchmark validity. The largest epistemic load is translation faithfulness and the overlap between the translation model and an evaluated model.

free parameters (3)

answer tolerance = 1e-6
Hand-chosen numerical tolerance for marking outputs correct; affects reported accuracy but is not fitted to data.
execution timeout = 300 seconds
Hand-chosen sandbox time limit; models that run longer are marked as timeout failures, influencing accuracy.
memory limit = 32 GB
Hand-chosen sandbox memory cap; can induce memory-limit failures in large LP/MIP formulations.

axioms (5)

domain assumption Ground-truth answers in the five source English benchmarks are correct and numerical answers transfer to Japanese translations unchanged.
The paper does not re-solve or independently verify source answers; it inherits them in §2.1.
domain assumption The LLM translation (gpt-oss-120b) plus single-author post-editing preserves problem semantics, constraints, and numerical consistency for all 1,319 problems.
Translation methodology in §2.2; no inter-annotator agreement or back-translation check is reported.
domain assumption Accuracy on code-generation tasks with OR-Tools is a valid proxy for OR formulation ability.
The protocol equates solving the generated Python program with formulation quality in §2.3.
standard math McNemar's test on per-problem binary outcomes is valid for these data.
Applied per model–dataset pair in §3.2, but assumes outcomes are not systematically dependent across languages or problem structure.
domain assumption Single generations per problem are representative enough to compare language conditions.
No sampling temperature or number of runs is reported; McNemar is computed on one binary outcome per problem in §3.2.

pith-pipeline@v1.3.0-alltime-deepseek · 9367 in / 12210 out tokens · 107519 ms · 2026-08-01T19:57:46.758421+00:00 · methodology

0 comments

read the original abstract

We present JOR-Bench, a collection of five Japanese-language benchmarks for evaluating the ability of large language models (LLMs) to formulate and solve operations research (OR) problems. Each benchmark is a Japanese translation of an existing English benchmark: IndustryOR, MAMO Complex LP, NL4OPT, OptiBench, and OptMATH, covering 1,319 problems spanning linear programming, mixed-integer programming, non-linear programming, and combinatorial optimization. JOR-Bench is a solver-independent benchmark that can be used with any solver or programming language, and consists of pairs of Japanese problem statements and expected numerical answers. We evaluate seven LLMs, including multilingual general-purpose models and Japanese-specialized models, on both the original English and the new Japanese versions, and compare performance across languages. For the main evaluation, we standardize execution with the Python interface to OR-Tools to make model outputs comparable and reproducible with open-source software. Our results show that OR formulation ability is largely language-neutral for strong multilingual models; the overall average accuracy difference between English and Japanese is only $-0.3$ pp. Yet error analysis reveals subtle cross-lingual differences, including a pragmatic disambiguation failure in some domains that causes models to output decision-variable values instead of the objective value when the prompt is in Japanese.

Figures

Figures reproduced from arXiv: 2607.16777 by Yuu Jinnai.

**Figure 1.** Figure 1: Qwen3 vs. Swallow family gaps across datasets and languages. Hatches and color choices are designed to remain distinguishable in grayscale and for color-vision deficiencies. model capacity to reason about the optimization problem. 4.6 Qwen3 vs. Swallow: Effects of Japanese Continued Pretraining Because Swallow models are built on Qwen3 with additional Japanese-focused continued pretraining, we compare the … view at source ↗

**Figure 2.** Figure 2: MAMO backend sensitivity for the Swallow– [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

11 extracted references · 2 linked inside Pith

[1]

Antoine Dussolle, Andrea Carde \ n a D \'i az, Shota Sato, and Peter Devine. 2025. https://doi.org/10.18653/v1/2025.findings-naacl.344 M - IFE val: Multilingual instruction-following evaluation . In Findings of the Association for Computational Linguistics: NAACL 2025, pages 6176--6191, Albuquerque, New Mexico. Association for Computational Linguistics

doi:10.18653/v1/2025.findings-naacl.344 2025
[2]

Kazuki Fujii, Taishi Nakamura, Mengsay Loem, Hiroki Iida, Masanari Ohi, Kakeru Hattori, Hirai Shota, Sakae Mizuki, Rio Yokota, and Naoaki Okazaki. 2024. Continual pre-training for cross-lingual LLM adaptation: Enhancing japanese language capabilities. In First Conference on Language Modeling

2024
[3]

Hillier and Gerald J

Frederick S. Hillier and Gerald J. Lieberman. 2005. Introduction to operations research. McGraw-Hill College

2005
[4]

Chenyu Huang, Zhengyang Tang, Shixi Hu, Ruoqing Jiang, Xin Zheng, Dongdong Ge, Benyou Wang, and Zizhuo Wang. 2025 a . Orlm: A customizable framework in training large models for automated optimization modeling

2025
[5]

Xuhan Huang, Qingning Shen, Yan Hu, Anningzhe Gao, and Benyou Wang. 2025 b . https://doi.org/10.18653/v1/2025.findings-naacl.146 LLM s for mathematical modeling: Towards bridging the gap between natural and mathematical languages . In Findings of the Association for Computational Linguistics: NAACL 2025, pages 2678--2710, Albuquerque, New Mexico. Associat...

doi:10.18653/v1/2025.findings-naacl.146 2025
[6]

Yuu Jinnai. 2026. Cost of reasoning in non-english languages: A case study on japanese. arXiv preprint arXiv:2607.10114

Pith/arXiv arXiv 2026
[7]

Hongliang Lu, Zhonglin Xie, Yaoyu Wu, Can Ren, Yuxuan Chen, and Zaiwen Wen. 2025. https://openreview.net/forum?id=9P5e6iE4WK Opt MATH : A scalable bidirectional data synthesis framework for optimization modeling . In Forty-second International Conference on Machine Learning

2025
[8]

Youmi Ma, Sakae Mizuki, Kazuki Fujii, Taishi Nakamura, Masanari Ohi, Hinari Shimada, Taihei Shiotani, Koshiro Saito, Koki Maeda, Kakeru Hattori, Takumi Okamoto, Shigeki Ishida, Rio Yokota, Hiroya Takamura, and Naoaki Okazaki. 2025. Building instruction-tuning datasets from human-written instructions with open-weight large language models. In Second Confer...

2025
[9]

NVIDIA. 2025. Nvidia nemotron nano 2: An accurate and efficient hybrid mamba-transformer reasoning model. arXiv preprint arXiv:2508.14444

Pith/arXiv arXiv 2025
[10]

Rindranirina Ramamonjison, Timothy Yu, Raymond Li, Haley Li, Giuseppe Carenini, Bissan Ghaddar, Shiqi He, Mahdi Mostajabdaveh, Amin Banitalebi-Dehkordi, Zirui Zhou, and 1 others. 2023. Nl4opt competition: Formulating optimization problems based on their natural language descriptions. In NeurIPS 2022 Competition Track, pages 189--203. PMLR

2023
[11]

Zhicheng Yang, Yiwei Wang, Yinya Huang, Zhijiang Guo, Wei Shi, Xiongwei Han, Liang Feng, Linqi Song, Xiaodan Liang, and Jing Tang. 2025. https://openreview.net/forum?id=fsDZwS49uY Optibench meets resocratic: Measure and improve LLM s for optimization modeling . In The Thirteenth International Conference on Learning Representations

2025

[1] [1]

Antoine Dussolle, Andrea Carde \ n a D \'i az, Shota Sato, and Peter Devine. 2025. https://doi.org/10.18653/v1/2025.findings-naacl.344 M - IFE val: Multilingual instruction-following evaluation . In Findings of the Association for Computational Linguistics: NAACL 2025, pages 6176--6191, Albuquerque, New Mexico. Association for Computational Linguistics

doi:10.18653/v1/2025.findings-naacl.344 2025

[2] [2]

Kazuki Fujii, Taishi Nakamura, Mengsay Loem, Hiroki Iida, Masanari Ohi, Kakeru Hattori, Hirai Shota, Sakae Mizuki, Rio Yokota, and Naoaki Okazaki. 2024. Continual pre-training for cross-lingual LLM adaptation: Enhancing japanese language capabilities. In First Conference on Language Modeling

2024

[3] [3]

Hillier and Gerald J

Frederick S. Hillier and Gerald J. Lieberman. 2005. Introduction to operations research. McGraw-Hill College

2005

[4] [4]

Chenyu Huang, Zhengyang Tang, Shixi Hu, Ruoqing Jiang, Xin Zheng, Dongdong Ge, Benyou Wang, and Zizhuo Wang. 2025 a . Orlm: A customizable framework in training large models for automated optimization modeling

2025

[5] [5]

Xuhan Huang, Qingning Shen, Yan Hu, Anningzhe Gao, and Benyou Wang. 2025 b . https://doi.org/10.18653/v1/2025.findings-naacl.146 LLM s for mathematical modeling: Towards bridging the gap between natural and mathematical languages . In Findings of the Association for Computational Linguistics: NAACL 2025, pages 2678--2710, Albuquerque, New Mexico. Associat...

doi:10.18653/v1/2025.findings-naacl.146 2025

[6] [6]

Yuu Jinnai. 2026. Cost of reasoning in non-english languages: A case study on japanese. arXiv preprint arXiv:2607.10114

Pith/arXiv arXiv 2026

[7] [7]

Hongliang Lu, Zhonglin Xie, Yaoyu Wu, Can Ren, Yuxuan Chen, and Zaiwen Wen. 2025. https://openreview.net/forum?id=9P5e6iE4WK Opt MATH : A scalable bidirectional data synthesis framework for optimization modeling . In Forty-second International Conference on Machine Learning

2025

[8] [8]

Youmi Ma, Sakae Mizuki, Kazuki Fujii, Taishi Nakamura, Masanari Ohi, Hinari Shimada, Taihei Shiotani, Koshiro Saito, Koki Maeda, Kakeru Hattori, Takumi Okamoto, Shigeki Ishida, Rio Yokota, Hiroya Takamura, and Naoaki Okazaki. 2025. Building instruction-tuning datasets from human-written instructions with open-weight large language models. In Second Confer...

2025

[9] [9]

NVIDIA. 2025. Nvidia nemotron nano 2: An accurate and efficient hybrid mamba-transformer reasoning model. arXiv preprint arXiv:2508.14444

Pith/arXiv arXiv 2025

[10] [10]

Rindranirina Ramamonjison, Timothy Yu, Raymond Li, Haley Li, Giuseppe Carenini, Bissan Ghaddar, Shiqi He, Mahdi Mostajabdaveh, Amin Banitalebi-Dehkordi, Zirui Zhou, and 1 others. 2023. Nl4opt competition: Formulating optimization problems based on their natural language descriptions. In NeurIPS 2022 Competition Track, pages 189--203. PMLR

2023

[11] [11]

Zhicheng Yang, Yiwei Wang, Yinya Huang, Zhijiang Guo, Wei Shi, Xiongwei Han, Liang Feng, Linqi Song, Xiaodan Liang, and Jing Tang. 2025. https://openreview.net/forum?id=fsDZwS49uY Optibench meets resocratic: Measure and improve LLM s for optimization modeling . In The Thirteenth International Conference on Learning Representations

2025