Does Jax's jvp use the directional derivative "trick"?

[DonghanHe]

I have been thinking about using the Pearlmutter trick/directional derivative trick for Hessian calculation, and per the autodiff cookbook, we can do a forward mode + reverse mode AD for that exactly. The jvp(grad(f), x, v) feels exactly like the Pearlmutter/Directional derivative trick, is that true? Does that mean jvp is calculated by doing forward mode AD with respect to t for function \phi(t) = f(x+t v)?