diff --git a/src/accumulators/default.jl b/src/accumulators/default.jl
index 0cf5dd7a7..4206c347a 100644
--- a/src/accumulators/default.jl
+++ b/src/accumulators/default.jl
@@ -26,7 +26,7 @@ abstract type LogProbAccumulator{T<:Real} <: AbstractAccumulator end
 Create a new `LogProbAccumulator` accumulator with the log prior initialized to zero.
 """
 (::Type{AccType})() where {T<:Real,AccType<:LogProbAccumulator{T}} = AccType(zero(T))
-(::Type{AccType})() where {AccType<:LogProbAccumulator} = AccType{LogProbType}()
+(::Type{AccType})() where {AccType<:LogProbAccumulator} = AccType{LogProbType[]}()
 
 Base.copy(acc::LogProbAccumulator) = acc
 
@@ -175,7 +175,7 @@ function accumulate_observe!!(acc::LogLikelihoodAccumulator, right, left, vn, te
     return acclogp(acc, Distributions.loglikelihood(right, left))
 end
 
-function default_accumulators(::Type{FloatT}=LogProbType) where {FloatT}
+function default_accumulators(::Type{FloatT}=LogProbType[]) where {FloatT}
     return AccumulatorTuple(
         LogPriorAccumulator{FloatT}(),
         LogJacobianAccumulator{FloatT}(),
diff --git a/src/distribution_wrappers.jl b/src/distribution_wrappers.jl
index 33dd451ff..2668c9547 100644
--- a/src/distribution_wrappers.jl
+++ b/src/distribution_wrappers.jl
@@ -54,13 +54,13 @@ function Distributions.rand!(
     return Distributions.rand!(rng, d.dist, x)
 end
 function Distributions.logpdf(::NoDist{<:Univariate}, x::Real)
-    return zero(LogProbType)
+    return zero(LogProbType[])
 end
 function Distributions.logpdf(::NoDist{<:Multivariate}, x::AbstractVector{<:Real})
-    return zero(LogProbType)
+    return zero(LogProbType[])
 end
 function Distributions.logpdf(::NoDist{<:Matrixvariate}, x::AbstractMatrix{<:Real})
-    return zero(LogProbType)
+    return zero(LogProbType[])
 end
 
 for f in (
diff --git a/src/transformed_values.jl b/src/transformed_values.jl
index 1befbf435..ba7e0f3f3 100644
--- a/src/transformed_values.jl
+++ b/src/transformed_values.jl
@@ -362,7 +362,7 @@ function apply_transform_strategy(
         # vectorisation transform. However, sometimes that's not needed (e.g. when
         # evaluating with an OnlyAccsVarInfo). So we just return an UntransformedValue. If a
         # downstream function requires a VectorValue, it's on them to generate it.
-        (raw_value, UntransformedValue(raw_value), zero(LogProbType))
+        (raw_value, UntransformedValue(raw_value), zero(LogProbType[]))
     else
         error("unknown target transform $target")
     end
@@ -383,7 +383,7 @@ function apply_transform_strategy(
         (raw_value, linked_tv, logjac)
     elseif target isa Unlink
         # No need to transform further
-        (raw_value, tv, zero(LogProbType))
+        (raw_value, tv, zero(LogProbType[]))
     else
         error("unknown target transform $target")
     end
@@ -406,7 +406,7 @@ function apply_transform_strategy(
         (raw_value, linked_tv, logjac)
     elseif target isa Unlink
         # No need to transform further
-        (raw_value, tv, zero(LogProbType))
+        (raw_value, tv, zero(LogProbType[]))
     else
         error("unknown target transform $target")
     end
diff --git a/src/utils.jl b/src/utils.jl
index f89071349..68a15f62f 100644
--- a/src/utils.jl
+++ b/src/utils.jl
@@ -7,7 +7,7 @@ The type for all log probability variables.
 
 This is Float64 on 64-bit systems and Float32 on 32-bit systems.
 """
-const LogProbType = float(Real)
+const LogProbType = Ref(float(Real))
 
 """
     typed_identity(x)
@@ -42,7 +42,7 @@ behaviour.
 function typed_identity end
 @inline typed_identity(x) = x
 @inline Bijectors.with_logabsdet_jacobian(::typeof(typed_identity), x) =
-    (x, zero(LogProbType))
+    (x, zero(LogProbType[]))
 @inline Bijectors.inverse(::typeof(typed_identity)) = typed_identity
 
 """