Designing more efficient quantum circuits with Claude (arXiv 2610.22367)
A · Strong evidence Harvard, Anthropic and QuEra authors including Mikhail Lukin pair LLM proposals with an independent evaluator and report checkable Toffoli and qubit counts against published baselines on eight tasks.
Summary based on the full text·Original: about 45 minutes to read
In short
An arXiv preprint (15 October 2026) by Andi Gu (Harvard, Anthropic STEM Fellows), Alek Kemeny (Anthropic), Susanne Yelin and Mikhail Lukin (Harvard, QuEra). An LLM proposes quantum circuit designs as Python programs; an independent evaluator checks correctness and scores Toffoli and qubit cost. Across eight compilation tasks the discovered circuits beat published designs on seven, and on the eighth the search found a new reversible architecture for prime-field modular inversion.
Key numbers
- seven of the eighttasks where discovered circuits cut both Toffoli and qubit counts versus the state of the art
- 12Toffoli-qubit cost improvement factor for the Reed-Solomon DQI decoder
- 39improvement factor for the fixed-point Gaussian circuit, the largest in the table
- 4.0 million Toffolis and 935 qubitscomplete benchmark circuit built on the new reversible modular inversion
As reported by the source.
English·Web page·opens anthropic.com
What it reports
- Toffoli-qubit cost improvements over baselines: Reed-Solomon DQI decoder 12x, HHL rotation angle 32x, Coulomb potential 9.9x, fixed-point Gaussian 39x, Fe2S2 alias-select 4.1x, non-abelian lattice-gauge simulation 7.7x. The Hermitian DQI decoder drops from 1.0x10^8 to 9.3x10^5 Toffolis.
- Modular inversion: the search made divsteps, a classical algorithm previously impractical in reversible form, workable by decoding branch decisions from the current state instead of storing them. Full benchmark circuit: 4.0 million Toffolis and 935 qubits.
- Three recurring design principles: conditional routing instead of conditional arithmetic, recovering branch history during uncomputation, and a failure budget that drops rare costly inputs (stack overflow probability 2.4x10^-6 per inversion, saving 203 history qubits).
- The LLM identified the invariant that makes branch recovery work; the authors developed it into a proof.
How they got it
An evolutionary search: the LLM first proposes several algorithm families, each optimised in parallel with evaluator feedback, with strong candidates shared across families. Correctness is checked by simulating basis-state inputs, which reversible circuits allow cheaply.
Limits
- Designs and resource counts only; nothing was run on quantum hardware.
- Not peer reviewed at release.
- Ranking by Toffoli times qubit count is one choice; real hardware may weight depth, routing or error correction differently.
- Task boundaries and input distributions were hand-written by the authors.
Why it matters for companies
The working pattern is the point for most firms: when a design problem has a cheap, trustworthy automated checker, LLM search can find structures experts missed. That applies to chip layout, scheduling, compiler passes or pricing logic as much as quantum circuits. Firms with quantum roadmaps should also note that resource estimates for useful algorithms keep falling.
Summary based on the full text. Read the original for the full argument and data.Read the original on anthropic.com
- Tags
- AI for science