Google team uses AI agents and Lean proofs to cut quantum program costs up to 11-fold
Researchers at Google DeepMind and Google Quantum AI posted a preprint on arXiv (not yet peer reviewed) describing AI agents that propose quantum programs and prove them correct and within a stated resource budget using Lean, a proof-checking system. Cost was measured as logical qubits multiplied by T gates, and the team reported a 10.9-fold reduction for an inertial-confinement fusion stopping-power simulation (4,374 logical qubits, down from 5,650), 8.08-fold for a photochemistry model with 187 electrons and 47 nuclei, and 2.49-fold for a 63-electron, 58-orbital model of cytochrome P450 Compound I. The study also reported a more than 290-fold reduction for computing arcsine and up to a 1.41-fold cut in query rounds for parallel quantum search. Human researchers chose the problems and wrote the specifications, and the team says it plans to release its Lean library, specifications and verified solutions as open source.