# Google DeepMind agents write Lean-verified quantum programs beating prior baselines

> Google DeepMind and Google Quantum AI build a quantum program library in the Lean theorem prover, where teams of Gemini 4 agents write and prove programs.

- Status: official (From the company, authors or authority themselves, with a link to their own page.)
- Topic: Research
- Published: 2026-11-15T17:51:03+00:00
- Updated: 2026-11-15T18:10:24.732037+00:00
- Web page: https://ai-news.ikangai.com/story/google-deepmind-agents-write-lean-verified-quantum-programs-beating-prior-baseli/
- Seen in: DAIR.AI (Elvis Saravia)
- Change: research (Research)
- Applies to: Gemini 4, Lean theorem prover

## What changed

Teams of Gemini 4 agents use Lean proof checking, instead of simulation, as the reward signal when writing quantum programs, reaching up to 2.55x gains for parallel search and 290x for quantum arithmetic over prior baselines.

Google DeepMind and Google Quantum AI report that teams of Gemini 4 agents, working inside the Lean theorem prover, wrote quantum programs. The programs beat prior baselines by up to 2.55x for parallel search and 290x for quantum arithmetic. The arXiv listing for the paper, "Discovering formally verified quantum programs with AI," names Sergey Shirobokov and 24 co-authors, with Pushmeet Kohli as senior author.

The motivation is a bottleneck in how quantum algorithms are tested. Classical computers can only simulate quantum algorithms on small or restricted instances. A search loop that scores candidates by simulation therefore cannot reach the sizes where quantum algorithms would matter. The group's workaround replaces simulation with proof. Researchers write the correctness requirement and a cost target as a Lean theorem, which the authors call a formal specification. Agent teams propose programs, and the Lean kernel checks both correctness and cost. A separate comparator then re-verifies every accepted answer, so a program counts as a result only if it passes the proof check twice.

Agents work in teams that share a wiki of verified solutions, partial results and dead ends. They exchange ideas and critiques at scheduled conferences. According to the DAIR.AI roundup, ideas from one team can seed improvements in another.

The headline figures are the improvements over prior baselines: up to 2.55x for parallel search and 290x for quantum arithmetic. The abstract names the arithmetic result as arcsin, according to the Academy page. For simulations relevant to fusion, photochemistry and drug metabolism, the abstract reports gains of 11x, 8x and 2.5x.

The two versions of the paper give different parallel-search numbers. The DAIR.AI Academy page, dated Nov 6, reports up to 1.4x for parallel search. The arXiv abstract, revised on Nov 9 as version 2, reports up to 2.55x. The arXiv listing does not say what changed between the versions, and the revised paper currently claims 2.55x.

Proof checking matters here because the reward signal changes. A simulation-based search can only score programs at sizes a classical machine can run, so its feedback stops where simulation stops. A Lean proof of correctness and cost holds at any size the statement covers. The authors argue that this lets the agents explore algorithms that neither current classical nor quantum hardware can test directly.

## Why it matters

Formal proof checking can replace simulation as the reward signal, letting agents search for algorithms that current computers cannot test directly.

## Original source

- [(2611.09822) Discovering formally verified quantum programs with AI](https://arxiv.org/abs/2611.09822)
- [Discovering formally verified quantum programs with AI | DAIR.AI Academy | DAIR.AI Academy](https://academy.dair.ai/papers/discovering-formally-verified-quantum-programs-with-ai-2611.09822)

---

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.

Cite as: IKANGAI AI News, "Google DeepMind agents write Lean-verified quantum programs beating prior baselines", 2026-11-15, https://ai-news.ikangai.com/story/google-deepmind-agents-write-lean-verified-quantum-programs-beating-prior-baseli/
