HyperQ adds token-conditioned quantum residual branches to a frozen masked-diffusion language model, where a lightweight circuit hypernetwork emits per-token rotation angles and measurement axes for a shared sparse quantum circuit whose output is added back via a residual connection. Because the expectation values have an exact classical-simulation shortcut scaling linearly with qubit count, the authors trained circuits from 16 to 64 qubits inside a 1.1B-parameter backbone using only 20,000 fine-tuning examples versus 200,000 for classical baselines. At 64 qubits, HyperQ outperforms both the frozen backbone and a LoRA-adapted counterpart.
