My understanding from the bit of the thread before it derailed: It seems to give a 5-10% speedup (measured in ESS/gradient) for some models, with basically the same behaviour for others. Best speed up is for things like iid or lightly correlated normals (or ts with high dof). The cost of this is more divergences, which are caused by the worse resolution of the symplectic integrator due to a higher step size. Summary here
1 Like