# AutoOpt: A Dataset and a Unified Framework for Automating Optimization Problem Solving

**Authors:** Ankur Sinha, Shobhit Arora, Dhaval Pujara  
**arXiv:** [2510.21436](https://arxiv.org/abs/2510.21436)  
**Date:** 2025-10-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2510.21436v1/

## Key points

- Comprises over 11,000 handwritten and printed mathematical optimization models
- Includes single-objective, multi-objective, multi-level, and stochastic optimization problems with diverse complexities
- Labels in LaTeX representation provided for all images
- Modeling language representation for a subset of images
- Machine learning-based automated approach for solving optimization problems
- Users provide an image of the formulation to AutoOpt for efficient solving without human intervention
- Comprises three modules: M1 (Image_to_Text), M2 (Text_to_Text), and M3 (Optimization)
- MER task model (M1) outperforms ChatGPT, Gemini, and Nougat based on BLEU score metric
- Hybrid BOBD method (M3) demonstrates superior performance on complex test problems compared to traditional approaches
- Introduces an extensive dataset and innovative framework that streamlines automation of solving optimization problems using machine learning techniques

## Summary

This study introduces AutoOpt-11k, a groundbreaking image dataset comprising over 11,000 handwritten and printed mathematical optimization models. The dataset includes single-objective, multi-objective, multi-level, and stochastic optimization problems with diverse complexities such as non-linearity, non-convexity, non-differentiability, discontinuity, and high-dimensionality. Labels in LaTeX representation are provided for all images and modeling language representation for a subset of images. The dataset was created by 25 experts adhering to ethical data creation guidelines and verified through a two-phase process to minimize errors. The study also presents the AutoOpt framework - a machine learning-based automated approach for solving optimization problems. Users can simply provide an image of the formulation to AutoOpt which then efficiently solves it without requiring further human intervention. The framework comprises three modules: M1 (Image_to_Text), M2 (Text_to_Text), and M3 (Optimization). M1 utilizes deep learning for Mathematical Expression Recognition (MER) to generate LaTeX code from the optimization formulation in the image. M2 employs a fine-tuned LLM to produce PYOMO script (optimization modeling language) from LaTeX code. And finally, M3 uses a Bilevel Optimization based Decomposition (BOBD) method to solve the optimization formulation described in the PYOMO script. The study leverages the AutoOpt-11k dataset for training and testing deep learning models within the AutoOpt framework. Results show that the MER task model (M1) outperforms ChatGPT, Gemini, and Nougat based on BLEU score metric. Additionally, the hybrid BOBD method (M3) demonstrates superior performance on complex test problems compared to traditional approaches like interior-point algorithm and genetic algorithm. In conclusion, not only introduces an extensive dataset but also presents an innovative framework that streamlines the automation of solving optimization problems using machine learning techniques. This advancement holds significant potential for enhancing efficiency in various fields where mathematical programming formulations are prevalent.