A team proposes a new paradigm for compiling more efficient quantum circuits using artificial intelligence agents, provided that the critical subroutines of the circuit exhibit a classically verifiable structure. When these subroutines consist exclusively of reversible classical logic, phasing operations, and measurement-based uncomputation, their execution can be efficiently simulated for classical inputs. This property allows the approximate synthesis of the circuit to be certified through a form of fuzz testing, under suitable assumptions about the input distribution.

The pattern was recently exploited to improve quantum cryptanalysis circuits. Now, the authors highlight the flexibility of the approach with three notable applications: the computation of 1/√x, which bottlenecks real-space simulations of the Coulomb operator relevant to modelling inertial confinement fusion reactors; the decoding of Reed-Solomon codes, which hinders the decoded quantum interferometry algorithm when approximately solving the optimal polynomial intersection problem; and quantum look-up tables together with the controlled application of Givens rotations, which constrain the most performant quantum algorithms for simulating chemistry in second quantization.

The researchers present reference implementations of these algorithms and explain how they can be constrained and verified to seed a Kaggle competition designed to accelerate AI-based research into the optimization of their compilation.