arXiv preprint finds autoregressive drift limits exact quantum circuit synthesis beyond 26 gates

An arXiv preprint reports that autoregressive transformer models for quantum circuit synthesis suffer from a sharp degradation in exact functional equivalence as target circuit length grows. Evaluating a 44.8M-parameter encoder-decoder transformer on parameterized circuits (2-6 qubits) and Clifford+T circuits (3-6 qubits), the authors found exact-match rates dropped from 88% on circuits with nine or fewer gates to near zero beyond 26 gates. The failure is traced to autoregressive drift: early token errors cascade irrecoverably through left-to-right decoding. Inference-time candidate generation with equivalence verification raised exact-match rates from 7% to 22.5%, and scaling training data by 2.5x pushed them to 39.5%, but the length degradation persisted, even with more data, exact equivalence fell from 94% on short circuits to under 4% beyond 26 gates. The authors contrast this with parameterized circuits, where a hybrid approach using transformer structure plus classical angle optimization achieves median fidelity of 1.000, concluding that post-processing rescues approximate outputs but discrete exact-correctness requirements expose a fundamental limitation of autoregressive decoding.