A new paper describes an AI system that keeps notes while it hunts for equations, not just the equations themselves.
The paper, AutoSR: Automatic Symbolic Regression by Searching Research States (arXiv:2608.16876), introduces AutoSR, a system that pairs each candidate equation with the reasoning, computational checks, and review that led to it, bundling them into what the authors call a Research State. Two AI agents, a proposer and a reviewer, build these states while a search method called progressive-widening Monte Carlo tree search decides which lines of investigation get more computing time. At the end, the accumulated notes are synthesized into a report explaining why the winning equation was chosen. The paper reports that AutoSR recovered the correct relation on all nine test problems it tried, drawn from two existing benchmark suites, including three cp3-bench problems the authors say no previously published system has solved.
Most symbolic regression tools score equations mainly on how well they fit the data and how simple they look, which can reward a formula that matches noisy data but breaks down outside the range it was tested on. By keeping the reasoning trail instead of discarding it once an equation is scored, AutoSR is a bet that scientific credibility needs more than a good numerical fit.
That said, this is a result on nine benchmark problems chosen by the researchers themselves, not new physics discovered from scratch - the mixed record, three previously unsolved cp3-bench problems alongside six more standard LSR-Transform ones, measures a search method, not a scientific breakthrough.