Insider Brief
- Anthropic and Harvard researchers used Claude to design more efficient quantum circuits across eight tasks, reducing estimated computing resources for key routines used in quantum algorithms.
- One redesigned decoding circuit used roughly 12 times fewer costly quantum operations, while a new arithmetic design offered improved trade-offs between gate counts and qubit requirements.
- The results were evaluated through computational checks and resource estimates rather than demonstrations on quantum hardware, and practical benefits will depend on integration into complete algorithms and machines.
A team of researchers from Anthropic and Harvard used an artificial intelligence (AI) system to design quantum circuits that require fewer operations and qubits, potentially lowering the hardware demands of useful quantum computing.
The researchers, who posted their work on the preprint server arXiv, tested the approach on eight computational building blocks used in chemistry, optimization, physics simulation and number theory. Seven of those designs reduced both the number of costly quantum operations and the number of qubits compared with the study’s baselines. The eighth produced a new approach to arithmetic that offers better choices between computing time and memory.
The results concern circuit designs and resource estimates, rather than calculations demonstrated on a working quantum computer. The team nevertheless addressed a central problem for the industry that, in order for quantum computers to become more useful quantum, algorithms will need to improve computing alongside the development larger, more reliable machines.
According to the study, one redesigned decoding circuit required about 440,000 Toffoli gates, down from 5.1 million in the comparison design, while reducing its peak qubit requirement from 1,938 to 1,837. Toffoli gates are operations that are particularly costly in many proposed error-corrected quantum computers, making their number a useful measure of computational work.
The research also illustrates a potential role for AI in scientific development, according to an accompanying blog post on the work. Rather than asking a model simply to suggest ideas, the team gave it programming tools and an evaluator that checked proposed circuits and scored their resource requirements. The model could then revise its designs in response to concrete feedback.
Anthropic’s blog post identifies the model as Claude Mythos 5 and says its elliptic-curve arithmetic work took 53 hours, with human involvement limited to the initial instructions and occasional prompts to broaden its search.
Reducing the Cost of Quantum Computation
A quantum circuit is a sequence of operations applied to qubits, the units that store quantum information. Larger algorithms depend on smaller routines that perform arithmetic, retrieve data or calculate numerical functions.
Making those routines cheaper can reduce the requirements of the larger computation. But translating an efficient conventional algorithm into a quantum circuit introduces constraints that can consume much of the original advantage.
For the circuits examined in the study, calculations must preserve the information needed to reverse their operations. Temporary results cannot simply be overwritten as they would be in ordinary software. Instead, the circuit must clear its scratch work through a process called — ironically enough — uncomputation.
Keeping temporary results uses additional qubits and clearing them early can require repeating calculations later, increasing the number of gates, which are the computing steps for a quantum computer. Circuit designers must balance those competing costs.
The researchers’ AI search repeatedly found ways to recover information from the current state of a calculation rather than store a separate record of every earlier decision. Other improvements allowed different computational paths to share the same arithmetic operations.
A third strategy reduced resources reserved for rare, unusually demanding inputs. Where a task allowed a small probability of failure, the model could omit machinery needed only for those exceptional cases, the researchers write in the study.
These changes went beyond removing redundant gates from an existing design. They involved choosing algorithms, organizing data and planning how intermediate calculations would be reversed.
A Smaller Decoder and a New Arithmetic Design
One major result involved decoded quantum interferometry, or DQI, an algorithm that uses the wave-like behavior of quantum states to reinforce better answers and suppress poorer ones, making better solutions more likely when the result is measured.
The researchers examined a decoding routine used within DQI for a mathematical curve-fitting problem. This decoder identifies errors from a compact set of checks. Although it borrows techniques from error-correcting codes, its role here is inside the optimization algorithm, rather than correcting faults in the quantum hardware.
The AI-designed circuit replaced an approach based on the extended Euclidean algorithm with a method called Berlekamp–Massey. The replacement offered more efficient arithmetic, but a straightforward reversible implementation would have required substantial memory for intermediate information.
The new circuit avoided much of that storage by reconstructing information when needed.
According to the paper, one remaining storage requirement could reach 665 qubits in the worst case. The discovered design allocated 19 qubits instead, accepting failure on rare inputs that exceeded that capacity. A second stored entry was needed with a probability of about one in 10,000 under the specified input distribution.
Together, the changes reduced the decoder’s gate count by roughly a factor of 12.
Another result discussed by the researchers was focused on modular inversion, a type of division using numbers that cycle like the hours on a clock. It is a central operation in elliptic-curve calculations used in quantum algorithms relevant to cryptography.
The model developed a reversible implementation of divsteps, a conventional arithmetic method whose memory requirements had made it difficult to use efficiently in this quantum setting.
The key was recovering earlier computational decisions from patterns in the numbers that remained available. The study presents a mathematical proof explaining why that information could be recovered.
Recovering every decision was not always the cheapest option, however. The resulting circuit family combines reconstruction and storage in different proportions, allowing designers to choose between fewer qubits and fewer gates.
The study reports one complete benchmark circuit using about 4 million Toffoli gates and 935 qubits. Those figures describe an arithmetic component, not the resources required for an entire cryptographic attack.
Other tasks that showed savings included a circuit for calculating a fixed-point Gaussian function that reduced the product of gate count and qubit count by a factor of 39. A numerical routine associated with solving linear equations also improved that combined measure by a factor of 32, according to the paper.
What the Results Establish
The researchers began each search with a written task specification and a blank program. The model proposed several different algorithmic approaches, which were developed in parallel, with promising ideas shared between them.
An independent evaluator tested proposed circuits and ranked them by the number of Toffoli gates multiplied by peak qubit use.
This worked because the team selected a class of circuits whose behavior can be checked efficiently on conventional computers. General quantum circuits can be much harder to simulate, limiting how broadly the same process can be applied.
The comparisons also require care. For one of the seven tasks with lower gate and qubit counts, involving a Hermitian-code decoder, the researchers report that no previous reversible implementation existed in the literature. Its baseline was a straightforward reversible construction, rather than an established published quantum circuit.
The supplementary material describes extensive sampled tests and, for some components, exhaustive checks. Sampling can provide evidence of correctness and estimate rare failures, but it does not prove that every possible input will succeed.
The scoring system is also an approximation because actual execution costs depend on gate and qubit counts, hardware connections, operation scheduling, measurements and the system used to correct errors. A circuit with a better score under the study’s metric may not be the best choice for every machine.
Claude also showed limitations as a research tool. In early experiments, it favored established approaches and repeatedly checked apparent improvements, sometimes continuing after the results were already well supported, according to Anthropic’s post. That caution helped verify its work but could also discourage exploration of unfamiliar methods.
The researchers addressed this by requiring Claude to develop several distinct approaches in parallel and occasionally prompting it to broaden its search. According to the post, those interventions encouraged more ambitious designs, including the new approach to reversible modular inversion. Once Claude identified that approach, it developed, implemented and verified it autonomously.
From Individual Circuits to Larger Systems
The researchers propose extending the approach to coordinate decisions that are usually made separately, including algorithms, error-correcting codes and the physical architecture of a quantum computer.
That would require evaluators capable of tracking how changes in one part affect resource demands elsewhere. The study identifies designing an appropriate scoring system as a matter that still requires application and hardware judgment.
The immediate result is narrower but useful. AI-guided searches produced circuit designs that reduce the estimated cost of several important computational routines and, in modular inversion, developed a new reversible arithmetic structure.
The research team included Andi Gu, of Anthropic’s STEM Fellows Program and Harvard; Alek H. Kemeny, of Anthropic and Susanne F. Yelin and Mikhail D. Lukin, both of Harvard University. Lukin is also co-founder and Chief Scientist of QuEra Computing.
For a deeper, more technical dive, please review the paper on arXiv. It’s important to note that arXiv is a pre-print server, which allows researchers to receive quick feedback on their work. However, it is not — nor is this article, itself — official peer-review publications. Peer-review is an important step in the scientific process to verify results.
