PEARL Enables Interactive Optimization Modeling with LLMs.

Researchers have introduced PEARL, a new system designed to make optimization modeling more interactive and efficient by integrating large language models (LLMs) with actual solver feedback. Traditionally, LLMs attempt a one-shot generation of mathematical formulations from natural language, often without executing or iteratively refining the solution. PEARL mimics the real-world process of optimization modeling, which involves repeated cycles of solving, debugging, and revising. It uses Python execution and mathematical programming solvers directly within this loop. The system learns when to test partial models, how to interpret and revise based on solver diagnostics, and when the solution is finalized. Operating in a multi-turn, tool-integrated environment, PEARL leverages intermediate execution results, feasibility signals, and solution checks to continuously improve both the mathematical formulations and the solver code. This approach has shown substantial improvements in verified solve rates across diverse optimization benchmarks, with even smaller PEARL models outperforming much larger one-shot baselines.