AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-21
OptScientist: Automated Discovery of Optimization Laws
Many scientific phenomena are best understood as outcomes of underlying optimization problems, yet automated scientific discovery has largely focused on recovering equations that describe observed input-output patterns. OptScientist discovers compact symbolic optimization laws directly from data, combining a model that infers symbolic optimization problems by representing data as an optimization problem and extracting its symbolic form with an LLM agentic system that integrates theoretical priors from the literature, discovers candidate laws, and suggests new data to collect. It recovers ground-truth laws for over 80% of the 99 physics-derived problems in the OptFeynman benchmark while remaining robust to noisy context, and on human risky choice and central-bank monetary policy it finds compact laws with held-out predictive performance competitive with leading baselines using few parameters and far fewer tokens than LLM baselines.