diff --git a/cmake/VWFlags.cmake b/cmake/VWFlags.cmake
index 8a46c09197a..7c157df7516 100644
--- a/cmake/VWFlags.cmake
+++ b/cmake/VWFlags.cmake
@@ -16,6 +16,9 @@ if("${CMAKE_SYSTEM_PROCESSOR}" STREQUAL "x86_64")
   if(NOT "arm64" STREQUAL "${CMAKE_OSX_ARCHITECTURES}")
     # Use sse2 by default. Change to latest simd extensions such as avx512 on supported architecture.
     set(LINUX_X86_64_OPT_FLAGS -msse2 -mfpmath=sse)
+    if(UNIX AND NOT APPLE)
+      set(LINUX_X86_64_OPT_FLAGS ${LINUX_X86_64_OPT_FLAGS} -mavx2 -mfma)
+    endif()
   endif()
 endif()
 
diff --git a/vowpalwabbit/core/src/reductions/cb/details/large_action_space.h b/vowpalwabbit/core/src/reductions/cb/details/large_action_space.h
index a5d82684317..f421724b185 100644
--- a/vowpalwabbit/core/src/reductions/cb/details/large_action_space.h
+++ b/vowpalwabbit/core/src/reductions/cb/details/large_action_space.h
@@ -12,6 +12,11 @@
 #include "vw/core/v_array.h"
 #include "vw/core/vw_fwd.h"
 
+// Eigen explicit vectorization may not work well with AVX2 when using smaller MAX_ALIGN_BYTES.
+// For more info:
+// https://eigen.tuxfamily.org/dox/TopicPreprocessorDirectives.html#TopicPreprocessorDirectivesPerformance
+#define EIGEN_MAX_ALIGN_BYTES 32
+
 #include <Eigen/Dense>
 #include <Eigen/SparseCore>
 #include <iostream>