diff --git a/CMakeLists.txt b/CMakeLists.txt
index 79efc5a..e977216 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -27,7 +27,7 @@ macro(compile_options target_name)
     sanitizers(${target_name})
 endmacro()
 
-project(blt-gp VERSION 0.4.0)
+project(blt-gp VERSION 0.4.1)
 
 include(CTest)
 
diff --git a/examples/symbolic_regression.h b/examples/symbolic_regression.h
index a9b6ed9..3c206b1 100644
--- a/examples/symbolic_regression.h
+++ b/examples/symbolic_regression.h
@@ -194,10 +194,10 @@ namespace blt::gp::example
 			// TODO: make stats helper
 			const auto& stats = program.get_population_stats();
 			BLT_INFO("Stats:");
-			BLT_INFO("Average fitness: %lf", stats.average_fitness.load());
-			BLT_INFO("Best fitness: %lf", stats.best_fitness.load());
-			BLT_INFO("Worst fitness: %lf", stats.worst_fitness.load());
-			BLT_INFO("Overall fitness: %lf", stats.overall_fitness.load());
+			BLT_INFO("Average fitness: {:0.6f}", stats.average_fitness.load());
+			BLT_INFO("Best fitness: {:0.6f}", stats.best_fitness.load());
+			BLT_INFO("Worst fitness: {:0.6f}", stats.worst_fitness.load());
+			BLT_INFO("Overall fitness: {:0.6f}", stats.overall_fitness.load());
 		}
 
 		void execute()
diff --git a/include/blt/gp/program.h b/include/blt/gp/program.h
index f1daa28..d9c703f 100644
--- a/include/blt/gp/program.h
+++ b/include/blt/gp/program.h
@@ -462,7 +462,7 @@ namespace blt::gp
 					[this, &fitness_function, &crossover_selection, &mutation_selection, &reproduction_selection](const size_t id) {
 						thread_helper.barrier.wait();
 
-						multi_threaded_fitness_eval<FitnessFunc>()(fitness_function);
+						multi_threaded_fitness_eval<FitnessFunc>()(fitness_function, id);
 
 						if (thread_helper.next_gen_left > 0)
 						{
@@ -574,7 +574,7 @@ namespace blt::gp
 						const size_t id) {
 						thread_helper.barrier.wait();
 
-						multi_threaded_fitness_eval<FitnessFunc>()(fitness_function);
+						multi_threaded_fitness_eval<FitnessFunc>()(fitness_function, id);
 
 						if (thread_helper.next_gen_left > 0)
 						{
@@ -804,6 +804,10 @@ namespace blt::gp
 						current_stats.normalized_fitness.push_back(sum_of_prob + prob);
 						sum_of_prob += prob;
 					}
+					std::sort(current_pop.begin(), current_pop.end(), [](const auto& a, const auto& b)
+					{
+					    return a.fitness.adjusted_fitness > b.fitness.adjusted_fitness;
+					});
 					thread_helper.evaluation_left = 0;
 				}
 			};
@@ -812,9 +816,10 @@ namespace blt::gp
 		template <typename FitnessFunc>
 		auto multi_threaded_fitness_eval()
 		{
-			return [this](FitnessFunc& fitness_function) {
+			return [this](FitnessFunc& fitness_function, size_t thread_id) {
 				if (thread_helper.evaluation_left > 0)
 				{
+					thread_helper.barrier.wait();
 					while (thread_helper.evaluation_left > 0)
 					{
 						size_t size = 0;
@@ -828,6 +833,15 @@ namespace blt::gp
 																					std::memory_order::memory_order_relaxed));
 						perform_fitness_function(begin, end, fitness_function);
 					}
+					thread_helper.barrier.wait();
+					if (thread_id == 0)
+					{
+						std::sort(current_pop.begin(), current_pop.end(), [](const auto& a, const auto& b)
+						{
+							return a.fitness.adjusted_fitness > b.fitness.adjusted_fitness;
+						});
+					}
+					thread_helper.barrier.wait();
 				}
 			};
 		}
diff --git a/lib/blt b/lib/blt
index 9a05c86..0ebbc19 160000
--- a/lib/blt
+++ b/lib/blt
@@ -1 +1 @@
-Subproject commit 9a05c86b02c9c45c2b384c531007416148ec4b56
+Subproject commit 0ebbc198c5b80ca99e537460ef92fd806864fa3e
diff --git a/src/selection.cpp b/src/selection.cpp
index 14377e0..b2174df 100644
--- a/src/selection.cpp
+++ b/src/selection.cpp
@@ -20,12 +20,12 @@
 
 namespace blt::gp
 {
-    void select_best_t::pre_process(gp_program&, population_t& pop)
+    void select_best_t::pre_process(gp_program&, population_t&)
     {
-        std::sort(pop.begin(), pop.end(), [](const auto& a, const auto& b)
-        {
-            return a.fitness.adjusted_fitness > b.fitness.adjusted_fitness;
-        });
+        // std::sort(pop.begin(), pop.end(), [](const auto& a, const auto& b)
+        // {
+        //     return a.fitness.adjusted_fitness > b.fitness.adjusted_fitness;
+        // });
         index = 0;
     }
 
@@ -35,19 +35,19 @@ namespace blt::gp
         return pop.get_individuals()[index.fetch_add(1, std::memory_order_relaxed) % size].tree;
     }
 
-    void select_worst_t::pre_process(gp_program&, population_t& pop)
+    void select_worst_t::pre_process(gp_program&, population_t&)
     {
-        std::sort(pop.begin(), pop.end(), [](const auto& a, const auto& b)
-        {
-            return a.fitness.adjusted_fitness < b.fitness.adjusted_fitness;
-        });
+        // std::sort(pop.begin(), pop.end(), [](const auto& a, const auto& b)
+        // {
+        //     return a.fitness.adjusted_fitness < b.fitness.adjusted_fitness;
+        // });
         index = 0;
     }
 
     const tree_t& select_worst_t::select(gp_program&, const population_t& pop)
     {
         const auto size = pop.get_individuals().size();
-        return pop.get_individuals()[index.fetch_add(1, std::memory_order_relaxed) % size].tree;
+        return pop.get_individuals()[(size - 1) - (index.fetch_add(1, std::memory_order_relaxed) % size)].tree;
     }
 
     const tree_t& select_random_t::select(gp_program& program, const population_t& pop)