Google teams of LLM agents discover quantum programs with Lean-checked proofs, cutting one routine's cost 290-fold
The Google DeepMind and Google Quantum AI preprint reports machine-checked cost cuts of about 11x, 8x and 2.5x for fusion, photochemistry and drug-metabolism simulations; the programs target fault-tolerant machines that do not exist yet.
Why it matters
Large quantum algorithms cannot be tested on today's classical or quantum computers, so AI search that relies on numerical checks stops at small instances. The authors built a quantum program library in the Lean theorem prover, let teams of LLM agents propose programs, and had an independent verifier check every proof of correctness and cost against a frozen specification. Agent teams shared results through a wiki and scheduled virtual conferences. Watch whether the resource savings hold up in independent review and whether the Lean library is released for others to use.
Key facts
- arcsin routine: cost cut by 290x
- Parallel Grover search: up to 1.4x better than prior baselines
- Fusion, photochemistry, drug-metabolism simulations: 11x, 8x, 2.5x
- Proofs of correctness and cost checked in the Lean theorem prover
- Authors from Google DeepMind and Google Quantum AI
- arXiv 2611.09822, submitted 6 Nov 2026