Classically simulated quantum circuits raise a frozen 1.1B model's scores
single source· 1 articles · confidence: medium · first seen 2026-09-20 20:00 UTC
What this means for you
Nothing to build on yet: no code, weights or serving date, and the circuits were simulated classically, so no quantum hardware is involved. Treat the 47.65-to-54.30 spread as one unreviewed preprint on unnamed benchmarks. It is worth reading if you weigh cheap fine-tuning against adding new branches — but there is nothing to migrate to.
HyperQ, an arXiv preprint, freezes a 1.1-billion-parameter language model, one that fills in blanked text rather than predicting left to right, and trains only added branches. Each branch emits a token-specific quantum circuit and adds its measured values back. Those expectation values have an exact classical expression costing linearly in qubit count, so the circuits run on ordinary hardware; 16 to 64 qubits were used. The average benchmark score rises from 47.65 to 54.30 as circuits widen, and at 64 qubits beats the frozen backbone by 4.71 points and a low-rank-adapted baseline by 3.67. The paper gives no evaluation dates or benchmark names.
Key facts
- ·HyperQ adds token-conditioned quantum residual branches to a frozen 1.1-billion-parameter masked-diffusion language model; only the added branches are trained. source
- ·Circuits from 16 to 64 qubits were trained inside that 1.1-billion-parameter backbone. source
- ·The paper reports average downstream benchmark score rising from 47.65 to 54.30 as circuit width increases, without naming the benchmarks or giving evaluation dates. source
- ·At 64 qubits, HyperQ exceeds the frozen backbone by 4.71 points and a low-rank-adapted counterpart by 3.67 points. source
- ·HyperQ was fine-tuned on 20,000 prompt-response pairs, against 200,000 for the classical baselines. source
- ·The circuits' required expectation values have an exact classical expression whose evaluation cost grows linearly with qubit count. source
What the sources say
- Hugging Face Daily Papers (research) — Single arXiv preprint describing classically simulated quantum circuit branches added to a frozen language model.
Sources
The original reporting. Follow these — they did the work.
- Hugging Face Daily PapersCircuit Hypernetworks for Quantum-Augmented Diffusion Language Models2026-09-20