OptSkills: Learning Generalizable Optimization Skills from Problem Archetypes via Cluster-Based Distillation

OptSkills, a new AI system, advances automated optimization problem-solving by clustering problems by underlying mathematical archetypes rather than surface narratives, achieving 68.27% accuracy on diverse benchmarks and outperforming DeepSeek-V3.2-Thinking on large-scale problems. The system uses skill distillation and trajectory learning to improve generalization across both known and novel problem types.

OptSkills addresses a critical limitation in current LLM-based optimization systems: their brittleness when problem narratives shift while underlying mathematical structures remain similar. Traditional approaches treat each optimization problem as a unique case, requiring expensive retraining. This archetype-centric approach mirrors how human mathematicians recognize problem families—a linear programming problem remains structurally identical whether it optimizes shipping routes or manufacturing schedules.

The technical innovation centers on three generalization layers. First, clustering by mathematical archetype filters out narrative noise, allowing the system to recognize isomorphic problems. Second, within each cluster, the system explores multiple modeling paradigms and solver configurations, then distills successful approaches into reusable "skills"—essentially learned workflows that capture what works for that problem family. Third, the system continuously refines and expands its skill library with new problems, enabling adaptation to emerging problem types.

The benchmark results demonstrate meaningful progress. The 68.27% micro-averaged accuracy across diverse problems substantially advances the field, while the 26.91% performance on MIPLIB-NL's large-scale problems suggests the approach scales to real industrial complexity. Notably, the system improves out-of-distribution performance (72.79% on NLCO after training) without explicit OOD data, indicating genuine generalization rather than memorization.

For the optimization and operations research communities, this represents a shift toward more interpretable AI-assisted problem solving. The open-sourced code and skills democratize access to advanced optimization capabilities, potentially accelerating enterprise adoption of AI-driven decision systems across logistics, finance, and manufacturing.

• →OptSkills clusters optimization problems by mathematical archetype rather than narrative, improving robustness to superficial problem variations.
• →The system achieves 68.27% accuracy on diverse benchmarks and 26.91% on large-scale MIPLIB-NL problems, outperforming competing LLM approaches.
• →Skill distillation allows the system to reuse learned workflows across problem families, enabling both in-distribution and out-of-distribution generalization.
• →Out-of-distribution accuracy reaches 72.79% on NLCO benchmark after training on Nano-CO, demonstrating genuine transfer learning capabilities.
• →Open-source release of code and skills could accelerate enterprise adoption of AI-assisted optimization across logistics, finance, and manufacturing sectors.