Letting a language model steer the search turns out to be smarter than letting it write the answer.
Researchers tested three ways to pair language models with symbolic regression — a technique that hunts for mathematical equations fitting a dataset. In one setup, the model authors candidate equations; in another, it picks from options; in a third, it controls search parameters like which variables, operators, and depth the solver explores. That third role, implemented in a system called LLM-PySR, consistently outperformed the others across 74 equations from the AI-Feynman benchmark and seven additional complex formula-recovery tasks. Accuracy, complexity, stability, and computational cost all favored the controller setup. On a real-world battery dataset, LLM-PySR found a compact piecewise-linear relationship between early voltage-curve behavior and battery cycle life.
The finding reframes a common assumption in AI-assisted science: that the model's job is to generate or judge answers. Here, the symbolic solver does the heavy numerical lifting while the language model narrows where to look — which turns out to be a better use of what each system is actually good at. It also suggests that hybrid approaches, where neural and classical methods split responsibilities cleanly, may be underexplored relative to the current trend of asking models to do everything end-to-end.
Symbolic regression has existed for decades; what's new is using language models as a dial on the search rather than a replacement for it. Whether this division of labor scales to messier real-world science than benchmarks allow remains the obvious next question.