EvoOptiGraph: Weakness-Driven Coevolution via Graph-Based Structural Generation for Optimization Modeling

· Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Mathematics & Computational Sciences · Depth: Expert, quick

Summary

EvoOptiGraph, a novel framework, addresses two key challenges in automating optimization modeling from natural language with large language models (LLMs): the lack of structural diversity in training corpora and static data generation pipelines. This system represents each mixed-integer linear program (MILP) as an attributed bipartite graph, employing validity-preserving evolutionary operators to generate structurally diverse instances. These evolved graphs are then converted into solver code and natural language. Training proceeds in two stages: supervised fine-tuning (SFT) on an initial dataset, followed by reinforcement learning with verifiable rewards (RLVR), where graph-derived weakness signals guide the generation of new instances targeting the model's failures. This closed-loop coevolution of data and model significantly outperforms larger generalist models, agentic methods, and specialized baselines in accuracy, executability, and generalization across six public datasets.

Key takeaway

For AI Scientists and Machine Learning Engineers developing LLMs for optimization modeling, you should consider implementing data-model coevolution strategies. EvoOptiGraph demonstrates that guiding data generation with model weaknesses significantly boosts accuracy and generalization. Integrate graph-based structural generation and a two-stage SFT/RLVR training approach to overcome limitations of static training corpora and achieve superior performance on MILP tasks.

Key insights

Targeted data-model coevolution, driven by model weaknesses, significantly improves LLMs for optimization modeling.

Principles

Method

EvoOptiGraph represents MILPs as attributed bipartite graphs, applies evolutionary operators for diverse instance generation, then converts graphs to solver code and natural language for two-stage SFT and RLVR training.

In practice

Topics

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.