AutoSR Searches Persistent 'Research States' Instead of Equations, Recovering Three cp3-bench Relations No Published System Gets

arXiv 2608.16876·medium signalXBlueskyLinkedInCopy link

Symbolic regression normally keeps only a formula and a score, discarding the motivations and probes that would inform what to try next — and numerical fit plus syntactic complexity can't distinguish expressions that behave identically on the data but diverge outside it. AutoSR preserves a Research State coupling each candidate equation with its reasoning, computational evidence and independent review, developed by proposer-reviewer agents under progressive-widening MCTS that allocates compute across competing investigations. Across nine challenges from two benchmark suites it recovers algebraically equivalent relations in every case, including three cp3-bench problems no published system recovers and six structurally diverse LSR-Transform problems, and synthesizes a final report justifying the leading relation.