MM-OptBench: A Solver-Grounded Benchmark for Multimodal Optimization Modeling
Accepted at Conference on Neural Information Processing Systems (NeurIPS), 2026
Authors: Zhong Li, Qi Huang, Yuxuan Zhu, Mohammad Mohammadi Amiri, Niki van Stein, Thomas Bäck, Matthijs van Leeuwen, Zaiwen Wen, Lincen Yang.
MM-OptBench evaluates whether multimodal language models can turn textual and visual problem specifications into mathematical optimization models and executable solver code. The benchmark contains 780 solver-verified instances across six optimization families and three difficulty levels, enabling analysis of both visual data extraction and optimization modeling errors.