OptScientist studies a different target for automated discovery: the optimization problem that may generate observed behavior. Rather than only fit an equation between inputs and outputs, the framework tries to recover a compact symbolic account of what is being optimized and under which constraints.
A model inside a discovery process
The OptScientist paper combines a model that represents data as an optimization problem and extracts symbolic form with a language-model-based agentic system. The surrounding process incorporates theoretical priors from scientific literature, proposes candidate laws, abstracts mechanisms and suggests additional data to collect.
That combination separates producing a plausible formula from deciding which experiment or observation might distinguish it from alternatives. A compact law is useful only if its assumptions and predictions can be examined beyond the data used to propose it.
What the benchmark establishes
The authors introduce OptFeynman, a benchmark of 99 physics-derived optimization problems. They report recovering ground-truth laws for more than 80 percent of those problems and robustness to noisy contextual information, compared with the baselines included in their study.
The existence of a known ground truth makes that benchmark an informative recovery test. It is different from discovering the uniquely correct explanation of an open scientific phenomenon, where several models may fit the available observations.
Predictive performance is not causal proof
The paper also studies human risky choice and central-bank monetary policy. It reports compact optimization laws with held-out predictive performance competitive with leading baselines, using relatively few parameters and fewer tokens than language-model baselines.
Those results should be read as the authors’ predictive comparisons, not proof that the inferred objective is the actual causal mechanism behind every decision. For researchers, the useful next question is whether a proposed law survives new data and distinguishes itself from competing explanations. OptScientist provides a framework for generating and investigating those candidates; its broader scientific value depends on that subsequent scrutiny.
Comments