Anthropic and Harvard Use Claude to Discover Reversible Quantum Circuits With 17x Gate Reductions

The Three-Question Framework

What Happened?

On October 16, 2026, Anthropic and Harvard researchers reported that Claude, guided by an automated verification and scoring evaluator across 53 hours of autonomous computation, designed novel quantum circuits across eight fundamental computational primitives that surpass the best published human designs.

Why It Matters

Practical quantum computing is bottlenecked by physical qubit counts and Toffoli gate depths; by discovering a reversible modular inversion architecture that avoids storing intermediate branch histories, Claude reduced point-addition gate counts by 17x, moving high-value simulations like FeMoco nitrogenase closer to physical hardware feasibility.

What Could Happen Next

The research team is extending evaluator-guided LLM co-design to surface code compilation and error-correction decoders, while releasing the benchmark suite on arXiv for independent quantum compiler evaluation.

Executive Key Takeaways

In seven out of eight benchmark tasks, Claude discovered quantum circuits that simultaneously reduced both logical qubit allocations and Toffoli gate counts.

For prime-field modular inversion, Claude adapted the classical divsteps algorithm to quantum execution by proving mathematical identities in Lean that reconstruct branch history directly from output states.

Compared to the smallest previously known point-addition circuit, Claude's balanced design uses 12% more qubits while reducing total gate volume by a factor of 17.

In Decoded Quantum Interferometry (DQI), Claude replaced extended Euclidean decoders with Berlekamp-Massey architectures, cutting Reed-Solomon locator gate counts by 12x over Google Quantum AI's baseline.

Source: Anthropic research publication 'Designing more efficient quantum circuits with Claude' and accompanying arXiv preprint arXiv:2610.22367v1 ('Discovering state-of-the-art quantum circuits with large language models') by Andi Gu, Alek H. Kemeny, Susanne F. Yelin, and Mikhail D. Lukin, published October 15–16, 2026.

Unlike classical processors that discard intermediate scratch calculations during execution, fault-tolerant quantum algorithms must remain fully reversible to prevent phase decoherence and state corruption, requiring expensive uncomputation steps. When implementing modular inversion—the primary bottleneck in elliptic curve discrete logarithm arithmetic—standard implementations accumulated hundreds of auxiliary qubits to record branching decisions made during classical divsteps iterations.

Rather than storing complete branch histories, Claude developed an algebraic decoding rule that reconstructs intermediate branch states by analyzing numerical coordinate pairs after each step. To validate the approach at cryptographic scale, the model verified the formula through large-scale numerical simulations on 256-bit integers and formalized the underlying mathematical proofs inside the Lean interactive theorem prover, eliminating all stored branch memory registers.

By exploring trade-offs between uncomputing and memory caching, Claude constructed a complete Pareto family of point-addition circuits. In the leading configuration, the circuit requires 1,732 qubits and 1.84 million Toffoli gates, achieving a 17-fold reduction in gate count compared to Schrottenloher's 2026 baseline while adding only modest auxiliary qubit overhead.

For Decoded Quantum Interferometry (DQI), an algorithm candidate for superpolynomial quantum optimization speedups, Claude redesigned the Reed-Solomon error locator. Replacing the variable-time extended Euclidean approach with Berlekamp-Massey decoding and simulating 403 million decodings to establish that exceptional states require only 19 reserve qubits instead of theoretical worst-case reserves of 665 qubits, Claude reduced total decoding gates by a factor of 12 compared to Google Quantum AI's published design.

The Three-Question Framework

What Happened?

Why It Matters

What Could Happen Next

Executive Key Takeaways

Primary Sources & Citations