EvoOptiGraph: Coevolutionary Graph Generator for Optimization

EvoOptiGraph is a framework for automating optimization modeling from natural language that closes the loop between data generation and model learning. Submitted to arXiv on 25 Jun 2026 (arXiv:2606.26578), the method represents each mixed-integer linear program as an attributed bipartite graph and uses model weakness signals to evolve structurally diverse training instances.

What is EvoOptiGraph?

EvoOptiGraph is a weakness-driven coevolution framework that targets two limitations in current LLM-based optimization modeling: lack of structural diversity in training corpora and static, decoupled data generation pipelines. The system encodes MILPs as attributed bipartite graphs, applies validity-preserving evolutionary operators to those graphs, and compiles evolved graphs into solver code and natural language for training and verification.

The paper lists six authors: Qingcan Kang, Mingyang Liu, Xiaojin Fu, Shixiong Kai, Tao Zhong, and Mingxuan Yuan. The submission claims the approach produces structurally diverse instances and forms a closed loop where generated data and the model co-evolve guided by identified weaknesses.

How does the co-evolution pipeline work?

EvoOptiGraph converts MILP instances into attributed bipartite graphs, mutates those graphs with evolutionary operators that preserve validity, compiles the graphs deterministically into solver code and natural language, verifies correctness via back-translation, and uses a two-stage training procedure: supervised fine-tuning (SFT) on an initial dataset followed by reinforcement learning with verifiable rewards (RLVR) that steers generation toward model failures. The loop repeats as weakness signals from RLVR guide new instance generation.

More specifically, the framework applies evolutionary operators directly to graph-structured MILP representations to increase structural diversity. The evolved graphs are deterministically compiled into both solver-executable code and natural language descriptions, then verified through back-translation. Training first uses SFT on an initial dataset, then switches to RLVR where rewards are verifiable and where graph-derived weakness signals select or produce new instances that target the model's failure modes.

How well does it perform?

The authors report empirical results across six public datasets showing EvoOptiGraph "significantly outperforms larger generalist models, agentic methods, and specialized baselines" on three evaluation axes: accuracy, executability, and generalization. The paper frames those improvements as evidence that targeted data-model coevolution improves LLM performance on optimization modeling tasks.

The submission emphasizes two concrete design choices that enable those gains: the attributed bipartite graph representation of MILPs, and the reinforcement-learning phase with verifiable rewards (RLVR) that uses weakness signals to guide data generation.

Why it matters

EvoOptiGraph changes where effort is spent: instead of relying on static corpora or brute-force scaling of generalist models, it pushes structural diversity into the training distribution and ties generation to measurable model failures. That matters for optimization modeling because syntactic correctness alone does not ensure solver-executable code or cross-dataset generalization; the paper argues targeted instance generation and verifiable RL rewards can close that gap.

If the claimed improvements hold across settings, practitioners building LLMs for code-to-solver tasks or natural-language-to-MILP pipelines would gain a practical mechanism to surface and harden against failure modes without manual dataset curation.

What to watch

Watch for accompanying code and data releases linked from the arXiv entry and for independent replication on the six public datasets the authors used. Also look for papers or repositories that detail the specific evolutionary operators and the RLVR reward design, since those components are central to reproducing the closed-loop gains reported in arXiv:2606.26578.