Speaker
Description
We introduce Cayley Unitary Adapters (CUAs), a hardware-efficient way to inject trainable quantum parameters into a frozen large language model (LLM). Each adapter is a block-diagonal unitary whose blocks are parameterised by the Cayley transform of a skew-symmetric matrix; the parameters are trained entirely classically by knowledge distillation, and the trained two-qubit blocks are then executed on real superconducting hardware via amplitude encoding. On Llama 3.1 8B, a single adapter with ∼6,000 extra parameters improves WikiText perplexity by 1.43%, with end-to-end autoregressive generation validated on a 156-qubit IBM Quantum System Two processor and shown to be robust to hardware noise. A systematic study on SmolLM2 (135M) shows perplexity improving monotonically with block dimension, a sharp noise-driven phase transition at three qubits, and an operator-entanglement analysis separating the CUA from a depth-matched variational circuit. To our knowledge this is the first quantum enhancement of a production-scale, pre-trained LLM for autoregressive generation on real gate-based hardware.
| Topic | Hybrid quantum-classical approaches for learning and optimization |
|---|---|
| Additional Topic | Quantum variational circuits and their applications |