The Pulse
Claude Cuts Quantum Decoder Resource Cost 12-Fold
Claude developed quantum-circuit designs that reduced resource counts across several computational tasks, including a Reed–Solomon decoder with a 12-fold lower combined resource cost. The work also introduces a reversible approach to modula
AI.info Team · Written with AI ·
“In seven of the eight tasks, the discovered circuits significantly reduce both Toffoli and qubit counts relative to the state of the art.”
Andi Gu, Alek H. Kemeny, Susanne F. Yelin and Mikhail D. Lukin, the paper’s authors, write.
Anthropic announced the results on Oct. 16, a day after the research paper appeared on arXiv. The team used Claude Mythos 5 to develop circuits for eight computational tasks, from numerical calculations and chemistry to optimization and number theory. The work’s headline result is a new reversible quantum decoder whose Toffoli–qubit resource cost is 12 times lower than the prior design.
Claude searched beyond one circuit design
The project paired Claude’s proposals with an independent program that checked whether each circuit produced the required result and measured its resource use. Researchers gave the model plain-language task specifications, then let it propose and revise circuit designs using feedback from the evaluator. The paper describes an evolutionary search: distinct approaches were developed in parallel, and promising ideas could be shared across them.
The evaluator ranked circuits using the product of Toffoli count and peak qubit count. Toffoli gates are a costly operation in fault-tolerant quantum computing; the paper uses their count as a proxy for runtime, while qubit count measures how much quantum memory a design needs. The researchers say the framework works particularly well when a circuit’s correctness can be checked efficiently with classical computation.
The Reed–Solomon decoder drops to 440,000 Toffolis
One test focused on decoded quantum interferometry, an algorithm for certain optimization problems. Its Reed–Solomon decoding step identifies errors from a compact set of summaries called syndromes. The previous circuit used a synchronized version of the extended Euclidean algorithm, whose input-dependent steps had to be implemented as a longer fixed sequence of quantum operations.
Claude switched to Berlekamp–Massey decoding, then addressed the extra memory that a reversible circuit needs to undo temporary calculations. Rather than keeping a record of most intermediate values, the new design reconstructs them from the circuit’s current state. In the paper’s benchmark, the decoder’s Toffoli count falls from about 5.1 million to 440,000, while its peak qubit count declines from 1,938 to 1,837; together, those changes produce the 12-fold reduction in the paper’s combined resource measure.
The smaller memory budget carries a defined trade-off. A worst-case design would reserve 665 qubits for a stack of saved values, but the new circuit reserves 19 qubits for one value. Under the input distribution used for the task, a second value is needed about once in 10,000 runs; those rare cases exceed the allocation and are allowed to fail within the task’s stated error budget.
A 53-hour search found a new route for modular inversion
For elliptic-curve point addition, Claude developed a different approach to modular inversion, a costly arithmetic operation used in the calculation. The team reports that the search took 53 hours with minimal human input, limited to the initial specification and occasional prompts to explore more broadly. The circuit uses divsteps, a classical algorithm that had not previously yielded a practical reversible quantum implementation because the circuit would normally have to preserve the history of its conditional operations.
Claude derived a way to recover much of that history from the updated arithmetic state, then checked the result with numerical simulations and a proof formalized in Lean, a proof-assistant language. The paper presents this as a new reversible architecture, rather than a small reduction to an existing quantum circuit. It also reports a family of designs that trade qubit use against gate count, allowing different resource choices instead of a single optimum.
Better circuit estimates do not demonstrate quantum hardware
The findings are algorithm and circuit-design results, not experiments run on a quantum processor. The paper’s gains depend on its benchmark definitions, allowed error rates and resource-scoring method; the authors note that hardware constraints can change the relative importance of qubit count, gate depth, routing and measurements. A design that wins under one cost measure may not be the best fit for a particular machine.
The work does show how an evaluator can let a language model search beyond edits to a fixed circuit: it can propose a different algorithm, have the candidate checked, and use measured costs to guide revisions. The paper reports reductions in both Toffoli and qubit counts for seven tasks; its eighth task, modular inversion, introduces a new architecture and a changed resource trade-off. None of those results removes the hardware requirements for useful quantum computation, but the decoder benchmark’s measured drop—from 5.1 million Toffolis to 440,000—sets a specific target for future implementations.