Can Large Models Write Industrial-Grade Optimization Algorithms? MIT Proposes FrontierOR to Set an Exam for AI
Can large language models (LLMs) design industrial-grade optimization algorithms? MIT researchers introduced FrontierOR, a benchmark evaluating LLMs on their ability to design scalable, high-quality algorithms for complex, large-scale optimization problems—going beyond simple modeling or solver calls. The benchmark, constructed from 180 real-world problems published in OR journals (1992-2025), assesses models in one-shot algorithm generation and self-evolution settings. Key findings show top models achieve high code execution rates (~0.98), but struggle to maintain feasibility and near-optimal solution quality on hard instances. Models like Claude Opus 4.6 exhibit more diverse algorithm design (e.g., decomposition, heuristics, hybrids), correlating with better performance. Self-evolution frameworks (e.g., CORAL) significantly boost results, raising the quality-time efficiency metric from 0.15 to 0.50 on the hardest tasks by iteratively refining algorithms. The study highlights a shift in failure modes from basic modeling errors to deeper challenges in heuristic search and structural exploitation. FrontierOR points toward future AI-driven optimization systems where LLMs act as algorithm designers, dynamically composing strategies and learning from feedback for applications in supply chain, energy, and transportation.
marsbit07/10 09:08
