A reinforcement-learning method reported substantially smaller errors than conventional fixed-step Trotterization in a simulated long-time quantum evolution, while carrying the same learned policy into larger systems. The method, called RL-Trotter, chose the next simulation step from signals about the system’s energy and energy variance, rather than using one step size throughout.
The result comes from a preprint based on classical numerical simulations of the mixed-field Ising benchmark described in the study. The method was not demonstrated on a quantum processor, so the findings describe an algorithmic comparison in simulation rather than a hardware result.
A controller that watches conservation laws
The main benchmark followed a 16-site system after a sudden change, or quench, from an antiferromagnetic setting in which the x- and z-field values were both 0.3 to a paramagnetic setting with values of minus 1.7 and minus 1.3. The calculation therefore began in one regime and tracked the ensuing dynamics in another.
At each Trotter step, the reinforcement-learning agent received the current energy and energy variance, their fixed starting values, and the amount of evolution time already simulated. It then selected the size of the next step. Training used classical simulations, with an actor-critic network whose hidden layers contained 300, 200 and 100 neurons across 3,000 episodes.
The comparison focused on errors in energy density and energy-variance density, and on how far the simulation could evolve under set error tolerances. These were the study’s main diagnostics for judging the competing protocols.
Lower errors and a longer run
In the main noise-free comparison, RL-Trotter’s average energy-density error was about 1.19 times 10 to the minus 2 J, compared with 1.47 times 10 to the minus 1 J for conventional Trotterization. Its average energy-variance-density error was about 7.15 times 10 to the minus 2 J squared, versus 4.96 times 10 to the minus 1 J squared for the conventional method. The reported averages were approximate and were not accompanied by interval estimates.
The learned controller also outlasted ADA-Trotter, another adaptive approach, under tight error limits of 0.03 J for energy and 0.09 J squared for energy variance. ADA-Trotter often became stuck at its smallest allowed step and reached a total evolution time of 7.72 J to the minus 1, around 60% of the time reached by RL-Trotter in that comparison.
A separate reward comparison pointed in the same direction. Over the first 100 steps of the 16-site quench, the best fixed-step protocol earned a total reward of 49.52 at a step size of 0.14 J to the minus 1, while RL-Trotter earned 81.30.
The advantage persisted with noisy feedback
The researchers also added simulated measurement noise to the energy signals. In 50 independent evaluations, RL-Trotter reached an average evolution time of 11.04 J to the minus 1, with a standard deviation of 0.09 J to the minus 1, compared with 7.94 J to the minus 1 and a standard deviation of 1.42 J to the minus 1 for ADA-Trotter. At step 40, the reported energy errors were 3.84 times 10 to the minus 2 J, with a standard deviation of 2.16 times 10 to the minus 2 J, for RL-Trotter and 9.34 times 10 to the minus 2 J, with a standard deviation of 3.34 times 10 to the minus 2 J, for ADA-Trotter. The corresponding energy-variance errors were 1.27 times 10 to the minus 1 J squared and 2.63 times 10 to the minus 1 J squared, respectively, with the reported standard deviations.
The noise test used two feedback-noise parameters set to 0.03 after training at 0.1. The spread in the reported evolution time was smaller for RL-Trotter than for ADA-Trotter in this comparison.
A policy trained small reached larger systems
The learned policy was trained at system size 8 and then applied without further training at sizes 12, 16 and 20. At size 20, its dynamics agreed well with the exact calculation available for that system size, according to the reported comparison.
The same pattern appeared in a large-system calculation using the study’s matrix-product-state method. After training at size 8 with exact diagonalization, RL-Trotter was evaluated at size 100. Its average errors were 1.62 times 10 to the minus 3 J for energy, 7.07 times 10 to the minus 3 J squared for energy variance and 3.47 times 10 to the minus 3 for antiferromagnetic order, compared with 8.15 times 10 to the minus 3 J, 3.46 times 10 to the minus 2 J squared and 1.39 times 10 to the minus 2 for conventional Trotterization.
Transfer had clear boundaries
The method did not transfer equally well across all starting conditions. Training sampled pre-quench ground states with x-field values between 0.3 J and 0.6 J and z-field values between 0.4 J and 0.6 J, while evaluation covered x-field values from 0 to 1 and z-field values from 0 to 2. The strongest transfer was reported for antiferromagnetic-like initial states; performance was unreliable when the starting states had fundamentally different physical character.
The results therefore describe transfer within the reported mixed-field Ising tests, not a general guarantee for every starting state or model. The training was performed with classical simulations, and in-hardware training was described as a possible extension rather than a demonstrated experiment.
Because the work was based on simulations, it does not demonstrate performance on a quantum processor. The document is identified as arXiv:2608.20139 version 3, dated 8 October 2026, and remains a preprint.
Resources and support
Data, code and trained models are deposited in Zenodo under DOI 10.5281/zenodo.22037440. Automatic access is restricted, with credentials available to human researchers on request. The authors report support from the Quantum Science and Technology-National Science and Technology Major Project, the National Natural Science Foundation of China, Peking University high-performance computing, the European Union ERC QuSimCtrl grant, Deutsche Forschungsgemeinschaft FOR 5919 and the ct.qmat cluster of excellence. The simulations also used the MPIPKS high-performance computing cluster.
Paper data and sources
Original title: Reinforcement LearningtoHarness Approximation Errors for Long-Time QuantumSimulation
Authors: Yu-Bo Shi, Markus Heyl, Roderich Moessner et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text