HyperQ adds trainable quantum residual branches to a frozen 1.1B diffusion LM
jiqizhixin · x · 2026-11-05
Researchers from Université de Montréal, Mila and McGill introduce HyperQ: trainable, token-specific quantum residual branches added to a frozen 1.1B masked-diffusion language model, framed as architectural adaptation rather than a quantum speedup claim. A rank-16 hypernetwork predicts rotation angles, couplings and measurement axes per token, with circuit readouts feeding back into attention representations. The sparse ring+chord circuit family admits exact classical expectation values linear in qubit count, so training and benchmarks run on efficient classical simulation. Six-benchmark average rises from 49.59% to 54.30%, scaling from 16 to 64 qubits, with fixed degree 4 keeping readout cost O(n) versus O(n²) for all-to-all skeletons.
More from Research
- Zuse paper: decision models for 27 European languages hit 87.5% accuracy at 27B, lag English suites behind Jev and Kev — kalyan_kpl · 2026-11-05
- RoboColosseum launches a real-world robot benchmark: 30+ policies blind-tested on hundreds of robots across 150+ scenes — micoolcho · 2026-11-05
- Merge-and-split k-means heuristics escape Lloyd's local optima, shows Nielsen & Nock — FrnkNlsn · 2026-11-05
- 3D scanned-objects dataset scanned-objects-glb trends on Hugging Face — codepawl · 2026-11-05
- PyRUA-REAL: coding agents give robots recursive self-improvement in the real world — Qiang Wang · 2026-11-05
- Microsoft's HCU unifies programmatic, semantic, and visual channels for computer-use agents — MicrosoftResearch · 2026-11-05