One dangerous aspect there is that state.transition caches the logjoint and gradient computations. So when you use the @set macro here, you're going to also keep the cached log-joint and gradient computation, which will then be out of sync with the parameters.

If you then naively pass this transition somewhere, say, into the next step call, AHMC.jl will use the incorrect logjoint eval in the MH step.

IMO the safe way is to use the explicit constructor of PhasePoint I believe without passing in the cached values. This should result in receomputation of this.

just realized that model is not passed into setparams!!, I think for now, we only set the parameters, then when use the transition, logp should be recomputed. (we can also later introduce some like setlogp or compute_logp etc.) I'll add a comment in the code

Agreed with @torfjelde's concern. Can we add logdensitymodel to

• setparams!!(state, logdensitymodel, params)
• getparam(state, logdensitymodel).

where logdensitymodel follows the LogDensityProblem interface. This would allow us to recompute the model's log probability (i.e., recompute_logp) inside these setparam!! functions on demand, which Turing's new Gibbs sampler uses.