Many scientific phenomena are best understood as outcomes of underlying optimization problems, yet most automated discovery efforts have focused on recovering equations that fit observed input‑output patterns. OptScientist introduces a novel framework that directly extracts compact symbolic optimization laws from data. It consists of two core components:
- A model that first learns to represent data as an optimization problem and then extracts its symbolic form;
- An LLM‑driven agentic system that embeds the model in a discovery loop, incorporates theoretical priors from the literature, generates candidate laws, abstracts their mechanisms, and proposes new data to collect.
Evaluated on OptFeynman, a new benchmark of 99 physics‑derived optimization problems, OptScientist recovers ground‑truth laws for over 80% of the cases, remains robust to noisy contextual information, and substantially outperforms machine‑learning, symbolic, and LLM‑based discovery baselines. In studies of human risky choice and central‑bank monetary policy, the framework discovers compact symbolic laws that achieve held‑out predictive performance comparable to leading baselines while using far fewer parameters and tokens than LLM baselines.
Overall, OptScientist expands automated scientific discovery to uncover optimization laws that generate observed phenomena, enabling automated theory building across physical, biological, behavioral, and social sciences.
Review