Discovering state-of-the-art quantum circuits with large language models

AI involvement AI-autonomous

Humans wrote task specifications and the evaluator. OpenEvolve used Claude Mythos 5 to propose and refine circuit architectures from blank programs; authors describe no substantive human intervention in the searches.

The methods describe blank-program circuit searches without substantive human intervention; task specifications and evaluators are human inputs.

Importance Notable

Demonstrates substantive circuit-design improvements across explicit tasks by jointly changing algorithms, representations and uncomputation.

Technical verification Partial

Only the stated computational, numerical, conditional or component checks are documented for this scope; the full claimed endpoint has not received a documented complete rigorous check in the reviewed evidence. Resource counts concern the paper’s kickmix circuit model and fixed task specifications, not measured hardware speed or an end-to-end quantum advantage. Some tasks allow bounded error or rare failure. T×Q is the chosen objective; routing, depth, decoding and hardware assumptions can change practical rankings. The Hermitian baseline is not a previous state-of-the-art reversible circuit. The evaluator was produced by the authors and is not an outside researcher assessment. AIProofWatch has not executed the circuits or reproduced the resource counts.

Human understanding Full

Named human researchers explain or reconstruct the scoped central mechanism in the cited paper/account; credit for that explanation is separate from technical checking. Humans wrote task specifications and the evaluator. OpenEvolve used Claude Mythos 5 to propose and refine circuit architectures from blank programs; authors describe no substantive human intervention in the searches.