Vol. I · Quantum Technology Every story links its source Sunday, November 1, 2026
QuantumVend Subscribe
Research

Oct 17, 2026 · 2 min read

Anthropic and Harvard say an LLM found more efficient quantum circuits

Coverage on Oct 17, 2026 from Quantum Computing Report reported that Anthropic and Harvard said a large language model, running in an evolutionary loop with an independent evaluator, found quantum circuits that reduce Toffoli and qubit counts relative to the state of the art. The work, posted to arXiv (2610.22367) as 'Discovering state-of-the-art quantum circuits with large language models,' also reported a new reversible modular-inversion design. Anthropic's release claims circuit-cost reductions of up to 9,100×.

A quantum circuit diagram showing qubits passing through gate and measurement operations
Illustrative A quantum circuit diagram of qubits through unitary gates and measurements; a generic circuit, not Anthropic's or Harvard's specific discovered circuit Ahmedrabiah, via Wikimedia Commons CC BY-SA 4.0

Coverage on Oct 17, 2026 from Quantum Computing Report reported that Anthropic and Harvard said a large language model running in an evolutionary loop with an independent evaluator found more efficient quantum circuits. The work, posted to arXiv (2610.22367) as “Discovering state-of-the-art quantum circuits with large language models,” proposes an LLM that proposes circuit designs and an independent verifier that scores them, tested across numerical work, chemistry, optimization, simulation and number theory.

According to the report, Anthropic’s release said the approach cut circuit cost by up to 9,100× and found a new reversible modular-inversion design in a single run. The arXiv abstract says that in seven of eight tasks the LLM-found circuits “significantly reduce both Toffoli and qubit counts relative to the state of the art,” and that the eighth task uncovered a new reversible architecture for prime-field modular inversion.

What is checkable is the shape of the work as reported: a named method (LLM proposes circuits, independent evaluator verifies and scores them), a named scope (eight tasks across numerical work, chemistry, optimization, simulation and number theory), a named result (reduced Toffoli and qubit counts in seven of eight tasks), and a named side result (a new reversible modular-inversion design). The specific 9,100× reduction is from Anthropic’s release.

What the abstract I could verify did not establish is the 9,100× figure or any specific qubit counts — those come from Anthropic’s release rather than the peer-reviewed abstract, which lists only Harvard authors. The abstract does not name Anthropic or QuEra, and does not report the resource multipliers that the release does.

The honest read is that this is an AI-assisted-design story, not a quantum-computing result. An evolutionary loop where a language model proposes circuits and a verifier scores them is a plausible method for finding cheaper decompositions, and reducing Toffoli and qubit counts is a genuine gain. But the headline 9,100× figure is a self-reported claim from the release, not a number in the abstract: the checkable part is that the method and the eight-task scope are reported, and the actual resource savings are what independent benchmarking will show.

Sources