GPT-6 Astra writes variational quantum circuit angles directly: no optimizer, 95–100% of the VQE spectrum

A language model can write the gate angles of a variational quantum circuit directly, without an optimizer and without seeing any score, matching a classical optimizer where the weights have a meaning it can reason about. Previously such circuit angles were trained by a classical optimizer, each loss evaluation was a run of the quantum device, gradients vanish for generic circuits, and the optimization problem is NP-hard in general. With the Hamiltonian variational ansatz on the transverse-field Ising chain, the model's angles capture 95–100% of the spectrum between the highest and lowest eigenvalue, equal to the optimizer; depth-one QAOA matches the optimizer on all six graphs; for H2 the model derives the energy as a function of the single angle and returns its exact minimizer.

The angles the model writes are themselves a strong warm start, and an optimizer started from them reaches its target almost immediately. An optimizer started from random angles needs hundreds of loss evaluations to reach the same target value. In VQE and QAOA, L-BFGS-B started from the model's angles reached the target in a median of one loss evaluation, with the model's point already within tolerance in 61% of VQE and 64% of QAOA requests.

The method scales with the interpretability of the weight space rather than its size: when a fixed circuit is deepened until individual weights can no longer be reasoned about, the model's accuracy decays toward chance. On 3 vs 8 digit classification the model scores 0.96 at 18 weights, above an optimizer trained on the same ten examples, and falls to 0.70 at 54 weights while the optimizer improves monotonically with depth. On ten classes, fixed circuits with 30 and 54 weights are usable for the optimizer at 0.393 and 0.570, while the model stays at chance throughout, 0.07–0.14; its replies describe the weights as heuristic.

All circuits are at most eight qubits and are simulated exactly without noise, so reported numbers are noise-free expectations rather than sampled estimates; the conclusions are limited to the transverse-field Ising chain, the Heisenberg chain, H2, six eight-node MaxCut graphs and handwritten-digit classification.

The problems are well known, and the model has almost certainly seen the Ising model, QAOA on regular graphs and the digits data set during training, so part of what is measured is recall and the results do not predict performance on novel Hamiltonians or data sets; the optimizer baseline is a deliberately modest L-BFGS-B with finite differences and a few restarts, a reference rather than a tuned upper bound; the study reports a single model, and whether other frontier models share its strategies and failure modes remains open.