Zapata and UMD Launch ForShor: AI Quantum Designs Must Pass Lean Proofs
Zapata Quantum and the University of Maryland launched ForShor.ai on November 18, requiring AI-proposed circuit designs for Shor's algorithm to pass Lean 4 formal proofs before leaderboard credit.

Why it matters
- Zapata Quantum and the University of Maryland announced ForShor.ai on November 18, 2025.
- AI-proposed circuit designs must pass formal Lean 4 proofs before earning leaderboard credit.
- A preprint submitted November 16 by Suresh, Courtney, Tao, and Cao reports gate-cost scaling of O(n²⁺ᵉ).
- Peter Shor, the algorithm's creator and Zapata Scientific Advisory Board member, endorsed the verification-first approach.
- Zapata will use ForShor as a proving ground for verification methods in its Quantum Pilot platform.
Zapata Quantum and the University of Maryland launched ForShor.ai on November 18, an open challenge where AI-proposed quantum circuit designs earn leaderboard credit only after passing formal Lean 4 proofs. The project targets a specific, checkable problem: optimizing the arithmetic circuits inside Shor's algorithm, the quantum method whose ability to factor large integers underpins its threat to modern cryptography.
The stakes extend beyond one leaderboard. ForShor addresses a structural weakness in AI-driven science: producing a plausible algorithm or explanation is easier than establishing that it is correct. By placing the search for better circuit designs inside a framework with explicit mathematical constraints, the project tests whether AI agents and formal verification (mathematical proofs checked by software, not just tested on examples) can work together productively.
What does ForShor actually ask participants to do?
Participants use AI to propose improvements to a formally verified implementation of Shor's algorithm, submit solutions to a public leaderboard, and explore documentation that links the underlying mathematics to machine-checked proofs. Zapata says contributors may become co-authors of a subsequent research paper.
The challenge deliberately narrows the task. It does not ask an AI system to invent an entirely new quantum algorithm without constraints. Instead, it gives researchers a specific optimization problem with checkable conditions.
Shor's algorithm is best known for its cryptographic implications. Its quantum component finds a repeating structure, called an order, in modular arithmetic; classical mathematical steps then use that result to help factor a large integer. ForShor focuses on the design of the arithmetic circuits involved in that computation.
The University of Maryland's project description identifies fast Toom–Cook multiplication as a key building block. This approach breaks a large multiplication problem into smaller pieces, creating choices about how to organize the arithmetic. Those choices determine how many gates (operations in a quantum circuit) and logical qubits (units of quantum information at the algorithmic level) a design requires. Reducing either requirement improves a design's prospects, although translating it into physical hardware introduces further costs.
What does the verification framework check?
The technical foundation is a preprint submitted November 16 by Anirudh Suresh, Joshua M. Courtney, Runzhou Tao, and Yudong Cao. It describes a harness that accepts a design specification called a policy, which defines choices in a recursive circuit construction and can come from either a human researcher or an AI agent.
The framework covers three connected claims:
- The implementation's mathematical correctness.
- The asymptotic growth of its gate cost.
- The correctness of translating the verified algorithm into a more detailed circuit representation.
The reported gate-cost scaling is O(n²⁺ᵉ), where n describes input size and ε is positive. This expresses how cost grows as the problem gets larger. It is not a claim that a particular 2,048-bit instance can run with a small fixed number of gates.
The implementation uses Lean 4 formal proofs. In formal verification, researchers state the required behavior precisely and supply proofs that a proof assistant can check. This goes beyond testing a program on selected examples — though the guarantees remain tied to the definitions and assumptions being proved.
How do AI agents and the proof system divide the work?
One part of the design space concerns instruction tables for recursive multiplication. Different arrangements of additions, shifts, and related operations change a circuit's cost. A proposed table must satisfy the framework's admissibility conditions before it can inherit the relevant correctness guarantees.
This division of labor is the project's core idea. An agent can explore candidate arrangements or suggest optimizations, while an independently checkable contract determines whether the result is valid. Researchers can then study why a successful candidate works, rather than treating its leaderboard score as the only output.
The project's interactive proof documentation supports that inspection. It shows how arithmetic choices fit into the larger order-finding circuit, and it connects explanations and diagrams to the corresponding Lean code so readers can examine the argument at different levels of detail.
Peter Shor, the algorithm's creator and a member of Zapata's Scientific Advisory Board, endorsed the approach in a statement: "ForShor puts formal verification, with humans in the loop, at the center of that process."
Yudong Cao, Zapata's CTO and a co-author of the preprint, described the initiative as a white-box learning and research environment.
Why does this matter for Zapata's Quantum Pilot platform?
ForShor extends Zapata's collaboration with UMD and serves a commercial purpose. Zapata will use the challenge as a proving ground for verification methods in its Quantum Pilot software platform, which combines three components:
- Quantum Graph, which connects applications, algorithms, research, and hardware roadmaps.
- Quantum Engine, which handles algorithm selection, resource estimation, experiment design, and benchmarking.
- Quantum Assurance, which combines formal methods with expert review before results inform a decision.
Resource estimation — asking what an algorithm would actually require to execute — anchors the connection. A mathematically valid design can still demand more qubits or runtime than an available machine can provide. Keeping that analysis tied to the verified implementation helps researchers avoid a specific failure mode: proving one circuit correct while estimating the cost of a different one.
Zapata positions Quantum Pilot as a bridge between potential use cases and technical feasibility, including comparisons against classical computing. ForShor tests the verification workflows that platform depends on, under open scrutiny.
What ForShor does not claim
The project's contribution is an inspectable software and mathematics framework. It does not demonstrate that today's quantum processors can factor cryptographically significant numbers, and it does not eliminate the engineering demands of fault-tolerant computing.
Physical error rates, error correction, hardware connectivity, and execution time remain separate considerations. Likewise, a machine-checked proof establishes a stated mathematical result; making that result understandable to researchers still requires clear definitions, documentation, and review.
The broader research idea survives those limits. Let AI explore designs aggressively, then require its proposals to pass explicit, reproducible checks. ForShor gives that approach a concrete starting point: a defined quantum task, an open implementation, and a community that can inspect both the candidate designs and the reasoning behind them. Whether that model transfers from circuit optimization to other AI-assisted science is the question the leaderboard results will begin to answer.
Source: unite.ai
More from Marcus Bennett
Show full bio
Senior reporter covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
745 articles
Related articles
- Zero of 135 AI Research Submissions Met Experts' Bar in New Benchmark
- MIT Plugs GPT-5.6 Sol Into Quantum Lab, and It Calibrates Qubits Overnight
- PSI Says AI System Produced Lean-Checked Proof of 2008 Quantum Conjecture
- AI System Produces Machine-Checked Proof of 18-Year-Old Quantum Conjecture
- OpenAI and Anthropic Claim AI Models Refuted the Exponential Time Hypothesis