diff --git a/include/pybind11/cast.h b/include/pybind11/cast.h
index c518169c4..1010ae3e8 100644
--- a/include/pybind11/cast.h
+++ b/include/pybind11/cast.h
@@ -1802,6 +1802,10 @@ struct function_call {
     /// The `convert` value the arguments should be loaded with
     std::vector<bool> args_convert;
 
+    /// Extra references for the optional `py::args` and/or `py::kwargs` arguments (which, if
+    /// present, are also in `args` but without a reference).
+    object args_ref, kwargs_ref;
+
     /// The parent, if any
     handle parent;
 
diff --git a/include/pybind11/pybind11.h b/include/pybind11/pybind11.h
index c8bab7703..6cedcc262 100644
--- a/include/pybind11/pybind11.h
+++ b/include/pybind11/pybind11.h
@@ -576,8 +576,8 @@ protected:
                     continue; // Unconsumed kwargs, but no py::kwargs argument to accept them
 
                 // 4a. If we have a py::args argument, create a new tuple with leftovers
-                tuple extra_args;
                 if (func.has_args) {
+                    tuple extra_args;
                     if (args_to_copy == 0) {
                         // We didn't copy out any position arguments from the args_in tuple, so we
                         // can reuse it directly without copying:
@@ -594,6 +594,7 @@ protected:
                     }
                     call.args.push_back(extra_args);
                     call.args_convert.push_back(false);
+                    call.args_ref = std::move(extra_args);
                 }
 
                 // 4b. If we have a py::kwargs, pass on any remaining kwargs
@@ -602,6 +603,7 @@ protected:
                         kwargs = dict(); // If we didn't get one, send an empty one
                     call.args.push_back(kwargs);
                     call.args_convert.push_back(false);
+                    call.kwargs_ref = std::move(kwargs);
                 }
 
                 // 5. Put everything in a vector.  Not technically step 5, we've been building it