SAGE Framework Boosts LLM Optimization Modeling Accuracy
A group of researchers has unveiled SAGE, a new framework that significantly improves how large language models develop precise and efficient optimization programs. By clarifying the modeling approach during data creation and after training, SAGE leverages a solver-verified multi-strategy dataset and incorporates supervised fine-tuning with Segment-Weighted GRPO to train a student model. The overall reward system considers format adherence, accuracy, and solver performance. When tested across eight benchmarks, SAGE increased the average pass@1 rate from 72.7 to 80.3, outperforming the leading open-source model. It also detects a wider range of correct formulations and enhances component-level diversity at pass@16 by 19-29%. On a larger scale, SAGE produces constraint systems with 14.2% fewer constraints.
Key facts
- SAGE is a strategy-aware framework for optimization modeling with LLMs.
- It uses a solver-verified multi-strategy dataset.
- Training involves supervised fine-tuning and Segment-Weighted GRPO.
- Composite reward covers format compliance, correctness, and solver efficiency.
- Tested on eight benchmarks spanning synthetic and real-world settings.
- Improves average pass@1 from 72.7 to 80.3 over open-source baseline.
- Increases component-level diversity at pass@16 by 19-29%.
- Reduces constraint count by 14.2% at largest scale.
Entities
Institutions
- arXiv