EvoOptiGraph Coevolves LLMs and Data for Optimization Modeling.

Automating the creation of optimization models from natural language using large language models (LLMs) faces challenges due to limited structural diversity in training data and static data generation pipelines. To overcome this, researchers propose EvoOptiGraph, a novel framework where both the training data and the LLM itself co-evolve, driven by the model's identified weaknesses. EvoOptiGraph represents mixed-integer linear programs (MILPs) as attributed bipartite graphs. It then applies validity-preserving evolutionary operators to these graphs to generate a wide array of structurally diverse instances. These evolved graphs are converted into solver code and natural language descriptions through deterministic compilation and verified back-translation. The training process involves two stages: initial supervised fine-tuning (SFT) on a baseline dataset, followed by reinforcement learning with verifiable rewards (RLVR). During RLVR, signals indicating the model's failures guide the generation of new, targeted instances, creating a continuous feedback loop that updates the training distribution. This approach has shown significant improvements in accuracy, executability, and generalization compared to other models.