Circuit Hypernetworks for Quantum-Augmented Diffusion Language Models
HyperQ trains token-specific quantum residual branches inside a frozen 1.1B masked-diffusion language model.
The paper says each branch emits circuit coordinates from a token’s hidden state, executes the circuit, and adds measured values back through a residual connection. Its circuit hypernetwork generates rotation angles, coupling strengths, and measurement axes within a shared sparse structure. The authors report linear classical evaluation cost with qubit count, allowing 16- to 64-qubit circuits. Wider circuits raised the average benchmark score from 47.65 to 54.30, with the 64-qubit version beating the frozen backbone and a low-rank-adapted baseline while using fewer fine-tuning pairs. HF Daily Papers' note
The paper says each branch emits circuit coordinates from a token’s hidden state, executes the circuit, and adds measured values back through a residual connection. Its circuit hypernetwork generates rotation angles, coupling strengths, and measurement axes within a shared sparse structure. The authors report linear classical evaluation cost with qubit count, allowing 16- to 64-qubit circuits. Wider circuits raised the average benchmark score from 47.65 to 54.30, with the 64-qubit version beating the frozen backbone and a low-rank-adapted baseline while using fewer fine-tuning pairs. HF Daily Papers' note
score 4