Signal AutoSR System Automates Scientific Investigation Beyond Equation-Level Symbolic Regression

Summary

A preprint posted to arXiv's cs.AI category on August 17, 2026 introduces AutoSR, a fully automated symbolic regression system that searches persistent 'research states' rather than isolated candidate equations. The authors argue that finite, noisy data often yield numerically competitive expressions implying very different behavior outside the observed data range, making numerical fit and syntactic complexity alone insufficient measures of scientific credibility; existing symbolic regression systems typically retain only the final formula and score, losing the reasoning that motivated the search. AutoSR instead preserves each candidate equation's motivations, computational evidence, and independent review in a Research State, developed by proposer-reviewer agent pairs under a progressive-widening Monte Carlo tree search that allocates computation across competing lines of investigation, with the accumulated record synthesized into a final report explaining the leading relation. Across nine selected challenges from two benchmark suites, the authors report AutoSR recovers algebraically equivalent relations in every case, including three problems from the cp3-bench suite that no previously published system recovers and six structurally diverse problems from LSR-Transform.