diff --git a/ext/DynamicPPLMarginalLogDensitiesExt.jl b/ext/DynamicPPLMarginalLogDensitiesExt.jl
index 905ee168b..714a3be2d 100644
--- a/ext/DynamicPPLMarginalLogDensitiesExt.jl
+++ b/ext/DynamicPPLMarginalLogDensitiesExt.jl
@@ -14,6 +14,20 @@ function (lw::LogDensityFunctionWrapper)(x, _)
     return LogDensityProblems.logdensity(lw.ldf, x)
 end
 
+function MarginalLogDensities.optimize_marginal!(
+    mld::MarginalLogDensities.MarginalLogDensity{<:LogDensityFunctionWrapper}, p2
+)
+    w0 = mld.u[mld.iw]
+    # Rebuild the optimization problem so that the current non-marginalized
+    # parameters are used by Optimization.jl.
+    prob = MarginalLogDensities.Optimization.OptimizationProblem(mld.f_opt, w0, p2)
+    sol = MarginalLogDensities.Optimization.solve(prob, mld.method.solver)
+    wopt = sol.u::typeof(w0)
+    objective = sol.objective::eltype(w0)
+    mld.u[mld.iw] .= wopt
+    return wopt, objective
+end
+
 """
     DynamicPPL.marginalize(
         model::DynamicPPL.Model,