I only read the abstract from the most recent linked article [1], but it sounds like the results are statistically correct. For this field, that makes a lot of sense, rather than computing strictly “accurate” velocity/acceleration from finite differences (see [2] for the basics on this integration method / domain).

Edit (to make my comment clearer): they came up with a method to calculate the velocities they need for solving these equations with a new closed-form solution that doesn’t depend on the timestep size. This lets the author combine the new calculation with previous work to take big steps, yet still converge.

[1] https://www.tandfonline.com/doi/figure/10.1080/00268976.2019... which is actually also https://arxiv.org/abs/1909.04380

[2] https://en.m.wikipedia.org/wiki/Verlet_integration