diff --git a/Project.toml b/Project.toml
index 9bb2749e..8c38f156 100644
--- a/Project.toml
+++ b/Project.toml
@@ -3,7 +3,7 @@ uuid = "80f14c24-f653-4e6a-9b94-39d6b0f70001"
 keywords = ["markov chain monte carlo", "probabilistic programming"]
 license = "MIT"
 desc = "A lightweight interface for common MCMC methods."
-version = "5.6.1"
+version = "5.7.0"
 
 [deps]
 BangBang = "198e06fe-97b7-11e9-32a5-e1d131e6ad66"
diff --git a/src/interface.jl b/src/interface.jl
index b58ced99..e31324dc 100644
--- a/src/interface.jl
+++ b/src/interface.jl
@@ -19,6 +19,20 @@ then `reduce(chainscat, c)` is called.
 chainsstack(c) = c
 chainsstack(c::AbstractVector{<:AbstractChains}) = reduce(chainscat, c)
 
+"""
+    getadtype(s::AbstractSampler)
+    getadtype(m::AbstractModel, s::AbstractSampler)
+
+Specify the `ADTypes.AbstractADType` to be used when sampling from model `m` using sampler `s`.
+
+If the model is not relevant, then the implementation of AbstractSampler can
+directly overload the single-argument method `getadtype(s::AbstractSampler)`.
+
+By default, this returns `nothing`.
+"""
+getadtype(::AbstractSampler) = nothing
+getadtype(::AbstractModel, spl::AbstractSampler) = getadtype(spl)
+
 """
     bundle_samples(samples, model, sampler, state, chain_type[; kwargs...])