ResearchPod Summary
Automating the translation of natural language into formal optimization models (like Mixed-Integer Linear Programs, or MILPs) is a major challenge for LLMs. Current methods often suffer from static training datasets that lack structural diversity and fail to adapt to the specific weaknesses of the model. This paper introduces EvoOptiGraph, a framework designed to bridge the gap between data generation and model training through a weakness-driven, coevolutionary loop.
EvoOptiGraph represents optimization problems as attributed bipartite graphs, where nodes represent variables and constraints, and edges store coefficient information. This graph-based representation allows the authors to apply genetic operators—such as crossover and mutation—to generate structurally diverse and valid optimization instances.
The training process follows a closed-loop cycle:
Empirical evaluations across six public benchmarks (including NL4Opt and MAMO) demonstrate that EvoOptiGraph significantly outperforms larger generalist models, agentic methods, and specialized baselines. The results show that the coevolutionary loop effectively expands the training distribution, leading to better accuracy, higher executability of generated code, and improved generalization to structurally complex problems. The authors' ablation studies confirm that both the structural diversity provided by graph evolution and the targeted nature of the weakness-driven generation are critical for performance gains.
This work shifts the paradigm of LLM training for specialized tasks from static dataset curation to dynamic, model-aware data synthesis. By treating data generation as an active, iterative process that responds to model failures, EvoOptiGraph provides a scalable and effective way to improve LLM reasoning in domains where structural correctness is paramount.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.