OptScientist aims to discover the optimization laws behind data
An October preprint presents an AI workflow for inferring compact symbolic optimization problems from observations rather than only fitting input-output equations.
Looking for the optimization problem, not just the curve
Many mathematical models describe a relationship between inputs and outputs. A preprint submitted on October 20 takes a different route: it asks whether an AI system can recover a compact symbolic optimization problem that could have generated the observations. The authors call their framework OptScientist.
From data to a mathematical objective
The proposed workflow first trains a model to represent data as an optimization problem, then extracts a symbolic form. An accompanying language-model agent brings in theoretical priors from scientific literature, proposes candidate laws, groups them by mechanism and recommends informative new measurements. In this setting, an "optimization law" is a statement of what quantity is being optimized and under what constraints, rather than merely a formula fitted to a dataset.
The authors introduce OptFeynman, a benchmark of 99 physics-derived optimization problems, and report recovering the benchmark's ground-truth laws on more than 80% of them. They also test examples involving risky choice and central-bank monetary policy, where they report competitive held-out prediction with compact symbolic models. Those figures are the authors' results for their datasets, comparisons and evaluation protocol.
A research direction with important limits
Recovering a short expression is not the same as establishing a scientific explanation. Candidate optimization laws still need domain scrutiny, tests on new data and, where relevant, mathematical proof. The paper is an unreviewed preprint, and MathsAI has not independently reproduced its experiments. Its contribution is nevertheless notable for mathematics-and-AI work: it treats symbolic optimization itself as the object an AI system should search for.