OptMATH: A Scalable Bidirectional Data Synthesis Framework for Optimization Modeling

Hongliang Lu, Zhonglin Xie, Yaoyu Wu, Can Ren, Yuxuan Chen, Zaiwen Wen · 2025

Read OptMATH: A Scalable Bidirectional Data Synthesis Framework for Optimization Modeling by Hongliang Lu, Zhonglin Xie, Yaoyu Wu, Can Ren, Yuxuan Chen, Zaiwen Wen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-11102.txt.

Description

Hongliang Lu and co-authors manufacture training data for optimization modeling in a 2025 paper. Language models could turn plain problem stories into math formulations, but good paired examples are scarce. Their framework, OptMATH, builds such pairs in both directions at scale. It starts from seed problems that already carry formulations.

From each seed, the system generates fresh problem data at controlled difficulty. A back-translation step writes the story in natural language. Then a forward check remodels the story and rejects pairs that fail to match. Accepted pairs form the training set. The rejects, after filtering, become a hard new benchmark with much longer problems than older ones. Models from half a billion to thirty-two billion parameters, trained on the data, beat rivals across modeling tests. The authors release the data set openly.

Catalog record

CREATORHongliang Lu, Zhonglin Xie, Yaoyu Wu, Can Ren, Yuxuan Chen, Zaiwen Wen
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE73 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-11102

Browse the full library · About the Tower of Babel