An LLM's first guess got a quantum medical model unstuck, at 61% accuracy

A small language model proposed starting parameters for a quantum neural network, and the resulting training run reached 61.4% binary classification accuracy in 1.1 seconds, 160 times faster than random initialization, according to a new arXiv preprint.

The paper applies AdaInit, a 2024 method from Zhuang and Cunningham, to the DMR-IR public mammography dataset. Instead of starting a parameterized quantum circuit at random, the authors make one query to Qwen2.5-3B-Instruct, a 3-billion-parameter open-weight model, and use its output to seed the circuit. On the binary benchmark, the LLM-chosen starting point carried 14.6 times the gradient variance of random init (0.0095 versus 0.0006), enough to escape the early-training dead zone that variational quantum models typically stall in.

That dead zone has a name: barren plateau, a property of parameterized quantum landscapes where gradients shrink toward zero and the model stops learning before it starts. AdaInit treats the LLM as a one-shot prior; the language model proposes a region of parameter space where the loss surface is more curved, and gradient-based training takes over.

At 61.4% accuracy, the result sits barely above coin-flip on a public benchmark, run inside NVIDIA's CUDA-Q GPU-accelerated quantum simulator, not on a quantum computer and not on patient data. The authors flag independent reproduction on different LLMs, different datasets, and actual quantum hardware as next steps; the preprint does not.