diff --git a/lib/pls/CMakeLists.txt b/lib/pls/CMakeLists.txt
index f30cd59..4b3201d 100644
--- a/lib/pls/CMakeLists.txt
+++ b/lib/pls/CMakeLists.txt
@@ -34,8 +34,10 @@ add_library(pls STATIC
         include/pls/internal/scheduling/scheduler_impl.h
         include/pls/internal/scheduling/task.h src/internal/scheduling/task.cpp
         include/pls/internal/scheduling/scheduler_memory.h src/internal/scheduling/scheduler_memory.cpp
-        include/pls/internal/scheduling/lambda_task.h include/pls/internal/data_structures/deque.h)
+        include/pls/internal/scheduling/lambda_task.h include/pls/internal/data_structures/deque.h
 
+        #        include/pls/algorithms/parallel_scan.h include/pls/algorithms/parallel_scan_impl.h)
+        )
 # Add everything in `./include` to be in the include path of this project
 target_include_directories(pls
         PUBLIC
diff --git a/lib/pls/include/pls/algorithms/invoke_parallel_impl.h b/lib/pls/include/pls/algorithms/invoke_parallel_impl.h
index ef634bd..a25336e 100644
--- a/lib/pls/include/pls/algorithms/invoke_parallel_impl.h
+++ b/lib/pls/include/pls/algorithms/invoke_parallel_impl.h
@@ -11,27 +11,27 @@ namespace pls {
 namespace algorithm {
 
 template<typename Function1, typename Function2>
-void invoke_parallel(const Function1 &function1, const Function2 &function2) {
+void invoke_parallel(Function1 &&function1, Function2 &&function2) {
   using namespace ::pls::internal::scheduling;
 
-  auto sub_task_1 = lambda_task_by_reference<Function1>(function1);
-  auto sub_task_2 = lambda_task_by_reference<Function2>(function2);
+  using task_1_t = lambda_task_by_value<Function1>;
+  using task_2_t = lambda_task_by_value<Function2>;
 
-  scheduler::spawn_child(sub_task_2);
-  scheduler::spawn_child_and_wait(sub_task_1);
+  scheduler::spawn_child<task_2_t>(std::forward<Function2>(function2));
+  scheduler::spawn_child_and_wait<task_1_t>(std::forward<Function1>(function1));
 }
 
 template<typename Function1, typename Function2, typename Function3>
-void invoke_parallel(const Function1 &function1, const Function2 &function2, const Function3 &function3) {
+void invoke_parallel(Function1 &&function1, Function2 &&function2, Function3 &&function3) {
   using namespace ::pls::internal::scheduling;
 
-  auto sub_task_1 = lambda_task_by_reference<Function1>(function1);
-  auto sub_task_2 = lambda_task_by_reference<Function2>(function2);
-  auto sub_task_3 = lambda_task_by_reference<Function3>(function3);
+  using task_1_t = lambda_task_by_value<Function1>;
+  using task_2_t = lambda_task_by_value<Function2>;
+  using task_3_t = lambda_task_by_value<Function3>;
 
-  scheduler::spawn_child(sub_task_3);
-  scheduler::spawn_child(sub_task_2);
-  scheduler::spawn_child_and_wait(sub_task_1);
+  scheduler::spawn_child<task_3_t>(std::forward<Function3>(function3));
+  scheduler::spawn_child<task_2_t>(std::forward<Function2>(function2));
+  scheduler::spawn_child_and_wait<task_1_t>(std::forward<Function1>(function1));
 }
 
 }
diff --git a/lib/pls/include/pls/algorithms/parallel_for_impl.h b/lib/pls/include/pls/algorithms/parallel_for_impl.h
index b787b44..32f9aa5 100644
--- a/lib/pls/include/pls/algorithms/parallel_for_impl.h
+++ b/lib/pls/include/pls/algorithms/parallel_for_impl.h
@@ -25,14 +25,13 @@ void parallel_for(RandomIt first, RandomIt last, const Function &function) {
     // Cut in half recursively
     long middle_index = num_elements / 2;
 
-    auto body2 = [=] { parallel_for(first + middle_index, last, function); };
-    lambda_task_by_reference<decltype(body2)> second_half_task(body2);
-    scheduler::spawn_child(second_half_task);
-
-    auto body1 = [=] { parallel_for(first, first + middle_index, function); };
-    lambda_task_by_reference<decltype(body1)> first_half_task(body1);
-    scheduler::spawn_child(first_half_task);
-    scheduler::wait_for_all();
+    auto second_half_body = [=] { parallel_for(first + middle_index, last, function); };
+    using second_half_t = lambda_task_by_reference<decltype(second_half_body)>;
+    scheduler::spawn_child<second_half_t>(std::move(second_half_body));
+
+    auto first_half_body = [=] { parallel_for(first, first + middle_index, function); };
+    using first_half_t = lambda_task_by_reference<decltype(first_half_body)>;
+    scheduler::spawn_child_and_wait<first_half_t>(std::move(first_half_body));
   }
 }
 
diff --git a/lib/pls/include/pls/internal/data_structures/aligned_stack.h b/lib/pls/include/pls/internal/data_structures/aligned_stack.h
index 2e04702..f1ab8eb 100644
--- a/lib/pls/include/pls/internal/data_structures/aligned_stack.h
+++ b/lib/pls/include/pls/internal/data_structures/aligned_stack.h
@@ -40,8 +40,8 @@ class aligned_stack {
   aligned_stack(pointer_t memory_region, std::size_t size);
   aligned_stack(char *memory_region, std::size_t size);
 
-  template<typename T>
-  T *push(const T &object);
+  template<typename T, typename ...ARGS>
+  T *push(ARGS &&... args);
   template<typename T>
   void *push();
   template<typename T>
diff --git a/lib/pls/include/pls/internal/data_structures/aligned_stack_impl.h b/lib/pls/include/pls/internal/data_structures/aligned_stack_impl.h
index 849971a..ed7d25c 100644
--- a/lib/pls/include/pls/internal/data_structures/aligned_stack_impl.h
+++ b/lib/pls/include/pls/internal/data_structures/aligned_stack_impl.h
@@ -6,10 +6,10 @@ namespace pls {
 namespace internal {
 namespace data_structures {
 
-template<typename T>
-T *aligned_stack::push(const T &object) {
-  // Copy-Construct
-  return new(push < T > ())T(object);
+template<typename T, typename ...ARGS>
+T *aligned_stack::push(ARGS &&... args) {
+  // Perfect-Forward construct
+  return new(push < T > ())T(std::forward<ARGS>(args)...);
 }
 
 template<typename T>
diff --git a/lib/pls/include/pls/internal/data_structures/work_stealing_deque.h b/lib/pls/include/pls/internal/data_structures/work_stealing_deque.h
index 4c89b6b..caec50f 100644
--- a/lib/pls/include/pls/internal/data_structures/work_stealing_deque.h
+++ b/lib/pls/include/pls/internal/data_structures/work_stealing_deque.h
@@ -81,8 +81,8 @@ class work_stealing_deque {
                                                           tail_{other.tail_.load()},
                                                           previous_tail_{other.previous_tail_} {}
 
-  template<typename T>
-  T *push_tail(const T &new_item);
+  template<typename T, typename Function, typename ...ARGS>
+  T *push_tail(const Function &after_creation, ARGS &&... args);
   Item *pop_tail();
   Item *pop_head();
 
@@ -94,8 +94,8 @@ class work_stealing_deque {
   work_stealing_deque_item *item_at(offset_t offset);
   offset_t current_stack_offset();
 
-  template<typename T>
-  std::pair<work_stealing_deque_item, T> *allocate_item(const T &new_item);
+  template<typename T, typename ...ARGS>
+  std::pair<work_stealing_deque_item, T> *allocate_item(ARGS &&... args);
 };
 
 }
diff --git a/lib/pls/include/pls/internal/data_structures/work_stealing_deque_impl.h b/lib/pls/include/pls/internal/data_structures/work_stealing_deque_impl.h
index 95987a5..45cebdf 100644
--- a/lib/pls/include/pls/internal/data_structures/work_stealing_deque_impl.h
+++ b/lib/pls/include/pls/internal/data_structures/work_stealing_deque_impl.h
@@ -23,28 +23,30 @@ offset_t work_stealing_deque<Item>::current_stack_offset() {
 }
 
 template<typename Item>
-template<typename T>
-std::pair<work_stealing_deque_item, T> *work_stealing_deque<Item>::allocate_item(const T &new_item) {
+template<typename T, typename ...ARGS>
+std::pair<work_stealing_deque_item, T> *work_stealing_deque<Item>::allocate_item(ARGS &&... args) {
   // 'Union' type to push both on stack
   using pair_t = std::pair<work_stealing_deque_item, T>;
   // Allocate space on stack
   auto new_pair = reinterpret_cast<pair_t *>(stack_->push<pair_t>());
   // Initialize memory on stack
   new((void *) &(new_pair->first)) work_stealing_deque_item();
-  new((void *) &(new_pair->second)) T(new_item);
+  new((void *) &(new_pair->second)) T(std::forward<ARGS>(args)...);
 
   return new_pair;
 }
 
 template<typename Item>
-template<typename T>
-T *work_stealing_deque<Item>::push_tail(const T &new_item) {
+template<typename T, typename Function, typename ...ARGS>
+T *work_stealing_deque<Item>::push_tail(const Function &after_creation, ARGS &&... args) {
   static_assert(std::is_same<Item, T>::value || std::is_base_of<Item, T>::value,
                 "Must only push types of <Item> onto work_stealing_deque<Item>");
 
   offset_t local_tail = tail_;
 
-  auto new_pair = allocate_item(new_item);
+  auto new_pair = allocate_item<T>(std::forward<ARGS>(args)...);
+  after_creation(&(new_pair->second)); // callback for time after creation but before being visible to others
+
   // Prepare current tail to point to correct next items
   auto tail_deque_item = item_at(local_tail);
   tail_deque_item->set_data(&(new_pair->second));
diff --git a/lib/pls/include/pls/internal/scheduling/scheduler.h b/lib/pls/include/pls/internal/scheduling/scheduler.h
index fb9eb20..242b912 100644
--- a/lib/pls/include/pls/internal/scheduling/scheduler.h
+++ b/lib/pls/include/pls/internal/scheduling/scheduler.h
@@ -79,19 +79,21 @@ class scheduler {
    * Helper to spawn a child on the currently running task.
    *
    * @tparam T type of the new task
-   * @param sub_task the new task to be spawned
+   * @tparam ARGS Constructor argument types
+   * @param args constructor arguments
    */
-  template<typename T>
-  static void spawn_child(T &sub_task);
+  template<typename T, typename ...ARGS>
+  static void spawn_child(ARGS &&... args);
 
   /**
    * Helper to spawn a child on the currently running task and waiting for it (skipping over the task-deque).
    *
    * @tparam T type of the new task
-   * @param sub_task the new task to be spawned
+   * @tparam ARGS Constructor argument types
+   * @param args constructor arguments
    */
-  template<typename T>
-  static void spawn_child_and_wait(T &sub_task);
+  template<typename T, typename ...ARGS>
+  static void spawn_child_and_wait(ARGS &&... args);
 
   /**
    * Helper to wait for all children of the currently executing task.
diff --git a/lib/pls/include/pls/internal/scheduling/scheduler_impl.h b/lib/pls/include/pls/internal/scheduling/scheduler_impl.h
index 50f7b04..1eb8c95 100644
--- a/lib/pls/include/pls/internal/scheduling/scheduler_impl.h
+++ b/lib/pls/include/pls/internal/scheduling/scheduler_impl.h
@@ -40,14 +40,14 @@ void scheduler::perform_work(Function work_section) {
   }
 }
 
-template<typename T>
-void scheduler::spawn_child(T &sub_task) {
-  thread_state::get()->current_task_->spawn_child(sub_task);
+template<typename T, typename ...ARGS>
+void scheduler::spawn_child(ARGS &&... args) {
+  thread_state::get()->current_task_->spawn_child<T>(std::forward<ARGS>(args)...);
 }
 
-template<typename T>
-void scheduler::spawn_child_and_wait(T &sub_task) {
-  thread_state::get()->current_task_->spawn_child_and_wait(sub_task);
+template<typename T, typename ...ARGS>
+void scheduler::spawn_child_and_wait(ARGS &&... args) {
+  thread_state::get()->current_task_->spawn_child_and_wait<T>(std::forward<ARGS>(args)...);
 }
 
 }
diff --git a/lib/pls/include/pls/internal/scheduling/task.h b/lib/pls/include/pls/internal/scheduling/task.h
index 7bb4a4b..c29cd3d 100644
--- a/lib/pls/include/pls/internal/scheduling/task.h
+++ b/lib/pls/include/pls/internal/scheduling/task.h
@@ -24,48 +24,48 @@ class task {
   data_structures::deque<task>::state deque_state_;
 
  protected:
-  // TODO: Double Check with copy and move constructors, try to minimize overhead while keeping a clean API.
   explicit task();
-  task(const task &other);
 
   /**
    * Overwrite this with the actual behaviour of concrete tasks.
    */
   virtual void execute_internal() = 0;
 
-  template<typename T>
-  void spawn_child(T &&sub_task);
-  template<typename T>
-  void spawn_child_and_wait(T &&sub_task);
+  template<typename T, typename ...ARGS>
+  void spawn_child(ARGS &&... args);
+  template<typename T, typename ...ARGS>
+  void spawn_child_and_wait(ARGS &&... args);
   void wait_for_all();
 
  private:
   void execute();
 };
 
-template<typename T>
-void task::spawn_child(T &&sub_task) {
+template<typename T, typename ...ARGS>
+void task::spawn_child(ARGS &&... args) {
   PROFILE_FORK_JOIN_STEALING("spawn_child")
   static_assert(std::is_base_of<task, typename std::remove_reference<T>::type>::value, "Only pass task subclasses!");
 
   // Keep our refcount up to date
   ref_count_++;
 
-  // Assign forced values (for stack and parent management)
-  sub_task.parent_ = this;
-  sub_task.deque_state_ = thread_state::get()->deque_.save_state();
-
   // Push on our deque
-  const T const_task = sub_task;
-  thread_state::get()->deque_.push_tail(const_task);
+  auto deque_state = thread_state::get()->deque_.save_state();
+  thread_state::get()->deque_.push_tail<T>([this, deque_state](T *item) {
+    // Assign forced values (for stack and parent management)
+    item->parent_ = this;
+    item->deque_state_ = deque_state;
+  }, std::forward<ARGS>(args)...);
 }
 
-template<typename T>
-void task::spawn_child_and_wait(T &&sub_task) {
+template<typename T, typename ...ARGS>
+void task::spawn_child_and_wait(ARGS &&... args) {
   PROFILE_FORK_JOIN_STEALING("spawn_child_wait")
   static_assert(std::is_base_of<task, typename std::remove_reference<T>::type>::value, "Only pass task subclasses!");
 
   // Assign forced values (for stack and parent management)
+  // TODO: Move this after construction
+  T sub_task{std::forward<ARGS>(args)...};
   sub_task.parent_ = nullptr;
   sub_task.deque_state_ = thread_state::get()->deque_.save_state();
   PROFILE_END_BLOCK
diff --git a/lib/pls/src/internal/scheduling/task.cpp b/lib/pls/src/internal/scheduling/task.cpp
index 5a4b1f4..8b20667 100644
--- a/lib/pls/src/internal/scheduling/task.cpp
+++ b/lib/pls/src/internal/scheduling/task.cpp
@@ -13,11 +13,6 @@ task::task() :
     parent_{nullptr},
     deque_state_{0} {}
 
-task::task(const task &other) :
-    ref_count_{0},
-    parent_{other.parent_},
-    deque_state_{other.deque_state_} {}
-
 void task::execute() {
   PROFILE_WORK_BLOCK("execute task")
   auto last_executing = thread_state::get()->current_task_;