Pulse sequences for interacting spins

In simulation, a candidate pulse sequence beats a published one >10× and stays ahead under disorder.

>10× lower infidelity · sim.

The best robust pulse sequence (B1) over 49 experiments against the published yxx48 benchmark, and its performance as added site disorder grows, staying above yxx48. — Simulation · plots from the campaign's records (preprint Fig. 5)

15–27×lower infidelity vs yxx48, disorder ≤ 2J

23 / 23robustness tests where B1 beats yxx48

49experiments in about 16 hours

The problem

Disorder and spin–spin interactions limit the coherence of solid-state sensing spins. Pulse sequences can average both away, but no general analytic method builds them; the published yxx48 came from deep reinforcement learning.

What we did

Researchers froze the score on the benchmark's simulated 3-spin chain; promoted candidates also faced 23 robustness tests. The collaborative agent ran 49 experiments; at a plateau, one researcher instruction to consult the literature gave the largest gain.

Result

The campaign's B1 has 27× lower infidelity than yxx48 (15× at disorder twice the spin coupling) and wins all 23 robustness tests. The top scorer, from a separate reasoning model, falls below yxx48 at modest disorder.

Keep in mind

Simulation only, in the benchmark paper's spin-chain model; an experimental test is being planned. Disorder was outside the frozen score and was tested afterwards.

In the preprint

Sec. IV.D, “Pulse sequence design,” and Fig. 5 of arXiv:2610.17193 describe the campaign and its checks in more detail.

Fig. 5 of the preprint, full size — Fig. 5 of the preprint: (a) the best-scoring sequences of each experiment, with the best score so far and the best score among robust sequences; (b) robustness to random site disorder, which the frozen evaluator did not include (all simulation). Click to enlarge.

Data: Simulation (dipolar spin-chain model; deterministic frozen evaluator)
Models (as stated in the preprint): Claude Fable 5 + ChatGPT 5.6 Sol at Pro effort (additional analytic reasoning)
PhysKit version (as stated in the preprint): 0.6.0, an earlier version
Where: Sec. IV.D, Fig. 5

We list the versions the preprint reports so readers can trace each result to its tools, not as a recommendation: for new work, use the latest release.