Anthropic and academic collaborators report that Claude Mythos 5 helped redesign one quantum-circuit decoder, reducing its gate count by a factor of 12 in their comparison. The result comes from a software-checked preprint, not a run on a quantum computer. Its decoder is part of a proposed optimization algorithm, not a quantum chip’s error-correction system.
Seven of eight tasks had lower reported gate and qubit counts
The team gave Claude plain-language specifications for eight computational tasks used inside larger quantum algorithms. Claude proposed circuit designs as Python programs. A separate evaluator checked whether each design produced the required outputs and counted its Toffoli gates and peak qubits. A Toffoli-gate count is one measure of circuit work; qubits provide the quantum system’s working memory.
The paper reports that seven of the eight tasks used fewer Toffoli gates and fewer qubits than the prior designs selected for comparison. The eighth produced a new reversible architecture for modular inversion, with a different balance between gates and qubits. These are the authors’ comparisons in an arXiv preprint submitted Oct. 15.
Anthropic identifies the eighth comparison as elliptic-curve point addition, whose most expensive step is modular inversion. In its resource comparison, one point-addition design uses 12% more qubits than the smallest previous design but 17 times fewer gates.
A new decoding design cut gates by storing less history
One task involved decoded quantum interferometry, an algorithm proposed for an optimization problem called optimal polynomial intersection. The algorithm uses a Reed-Solomon decoder, which infers error locations from a compact set of values called syndromes. Here, “decoder” refers to a step inside that algorithm, not a quantum computer’s error-correction decoder.
The earlier circuit used the extended Euclidean algorithm. Its number and order of steps depend on the input, but a quantum circuit must use a fixed sequence of gates across inputs in superposition. The researchers say Claude shifted the design to Berlekamp-Massey decoding, whose regular sequence of updates fit that constraint better.
Reversibility adds another cost: a quantum circuit cannot simply erase temporary information when it is done. The new design reconstructs most of that history from the current register values and saves only a small remainder. The paper estimates that a second stack entry is needed about once in 10,000 inputs under the algorithm’s sampled distribution. The circuit reserves one 19-qubit entry instead of the 665 qubits needed for the worst-case stack, so it permits decoding to fail on those rare inputs. Against the prior construction cited in the paper, the authors report 12 times fewer gates for this decoder circuit.
The comparison does not measure a quantum speedup
The paper is an arXiv preprint and reports software checks and circuit-resource estimates. It does not test the designs on a physical quantum processor or establish that a real system could run them reliably or outperform a classical computer.
Sources
- The authors’ arXiv paper, “Discovering state-of-the-art quantum circuits with large language models”
- Anthropic’s explanation of the study
- Illustrative photo by IBM Research, via Wikimedia Commons, licensed under CC BY 2.0. The pictured IBM Quantum System One was not used in this study; the image is unmodified.
.jpg){kind=link}