OptMATH: A Scalable Bidirectional Data Synthesis Framework for Optimization Modeling
Authors: Hongliang Lu, Zhonglin Xie, Yaoyu Wu, Can Ren, Yuxuan Chen, Zaiwen Wen
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experiments, we demonstrate that models of various sizes (0.5B-32B parameters) trained on Opt MATH achieve superior results on multiple modeling benchmarks, thereby validating the effectiveness and scalability of our approach. |
| Researcher Affiliation | Academia | 1College of Engineering, Peking University 2Beijing International Center for Mathematical Research, Peking University 3School of Mathematics Science, Peking University. Correspondence to: Zaiwen Wen <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 Feedback-Driven Problem Data Generation Require: Target complexity range [Smin, Smax], time limits [Tmin, Tmax], instance generator G, feasibility threshold Ftarget, max iterations T Ensure: Configuration Θ such that for PDi G(Θ): S(PDi) [Smin, Smax] (complexity), τi Tmax (solving time), Pr(fi = feasible) Ftarget 1: Initialize parameters via LLM: Θ0 L(prompt IC(Smin, Smax, Tmin, Tmax)) 2: for t = 1 to T do 3: Generate N PDs: {PDi}N i=1 G(Θt 1) 4: Compute metrics: S(PDi) (Eq. 4), τi (solving time), fi (feasibility) 5: Aggregate statistics: St = 1 N P S(PDi), τt = 1 N P τi, Ft = 1 |
| Open Source Code | Yes | The Opt MATH dataset and related resources are available at https://github.com/optsuite/ Opt MATH. |
| Open Datasets | Yes | The Opt MATH dataset and related resources are available at https://github.com/optsuite/ Opt MATH. ... We evaluate our fine-tuned model on five benchmarks: NL4OPT(Ramamonjison et al., 2021), MAMO(Huang et al., 2024), Industry OR(Tang et al., 2024), Opti Bench(Yang et al., 2025) and our newly constructed Opt MATH-Bench. |
| Dataset Splits | No | The paper mentions 'Opt MATH-Train' as a training dataset and 'Opt MATH-Bench' as a benchmark, and refers to pre-existing test sets for other benchmarks (e.g., 'we selected the test set from this dataset' for NL4OPT). It also states MAMO is 'divided into two main components, Easy LP and Complex LP, containing 652 and 211 instances, respectively'. However, it does not provide specific train/validation/test percentages or counts for its own generated 'Opt MATH-Train' dataset. |
| Hardware Specification | No | The computational resources were supported by the Center for Intelligent Computing and Song-Shan Lake HPC Center (SSL-HPC) in Great Bay University, Dongguan, China. This statement provides general information about computational resources but lacks specific hardware details such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | Yes | MOSEK Ap S. MOSEK Optimization Software, 2025. Version 11.0.3. |
| Experiment Setup | Yes | We adopt a supervised fine-tuning (SFT) approach to enhance the Auto Formulator s modeling capabilities. Specifically, we employ the Lo RA algorithm (Hu et al., 2021) for efficient parameter-efficient fine-tuning... We select the Qwen2.5 series (0.5B 32B) as our base models (Yang et al., 2024), and the hyperparameters are generally set as follows: initial learning rate of 1e-4, 1 3 epochs, Lo RA rank of 32, Lo RA alpha of 32, and Lo RA dropout of 0.1. |