Quantum Circuit Generation via test-time learning with large language models

Adriano Macarone-Palmieri and Rosario Lo Franco propose a memory-augmented test-time optimization framework for large language models tackling black-box scientific design problems. The approach combines episodic memory, score-difference feedback, and restart-from-best sampling. Evaluated on quantum circuit synthesis, the framework achieves maximum Meyer–Wallach entanglement on 25-qubit circuits within 45 oracle calls, while a random hill-climbing baseline stalls below 0.29. The results establish quantum circuit synthesis as a benchmark for test-time optimization.