diff --git a/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon.c b/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon.c
index 36c20fd127c6130dd84c6060d755c1c784b6e50b..f264c5c05d74743bc9bffa69562254d4fcb0a0f2 100644
--- a/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon.c
+++ b/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon.c
@@ -142,17 +142,18 @@ void kai_run_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon(
                         // Load the 2 u4 values from source
                         const uint8_t dst_byte = src_row[(col_idx + kr_block_idx) / 2];
 
+                        // NOLINTBEGIN(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
                         // extract i8 values from the 2 u4 values
-                        const uint8_t first_value = (dst_byte & 0xF) - rhs_zero_point;
-                        const uint8_t second_value =
+                        const int8_t first_value = (dst_byte & 0xF) - rhs_zero_point;
+                        const int8_t second_value =
                             col_idx + kr_block_idx + 1 >= k ? 0 : (dst_byte >> 4) - rhs_zero_point;
 
                         // Add the i4 value to the row sum
                         sum += (int32_t)first_value + (int32_t)second_value;
 
                         // Truncate i8 to i4 and write to dst
-                        // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
                         dst_kr_block[kr_block_idx / 2] = (second_value << 4) | (first_value & 0xF);
+                        // NOLINTEND(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
                     }
 
                     // Go to the next kr block for this row in the nr rows
diff --git a/test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp b/test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp
index d9850f3d516cd052a5f2dbbde58db7c3ca76e059..37c37c39c601c8196fc7c8433b7cb5970dd65bc0 100644
--- a/test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp
+++ b/test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp
@@ -290,10 +290,9 @@ TEST_P(MatMulTest_f32_qai8dxp_qsi4cxp, EndToEnd_RHS_nxk_qsi4cx) {
     // Generates input data.
     const auto ref_lhs = fill_random<float>(M * K, seed + 0);
     const auto ref_biases = fill_random<float>(N, seed + 2);
-    std::uniform_real_distribution<float> dist(-10.0, 1.0);
 
+    std::uniform_real_distribution<float> dist(-10.0, 1.0);
     std::mt19937 rnd(seed + 1);
-
     const auto ref_rhs = fill_matrix_raw<float>(1, N * K, [&dist, &rnd](size_t, size_t) { return dist(rnd); });
 
     // Runs the reference implementation.
@@ -421,7 +420,11 @@ TEST_P(MatMulTest_f32_qai8dxp_qsi4cxp, EndToEnd_RHS_nxk_qsu4cx) {
 
     // Generates input data.
     const auto ref_lhs = fill_random<float>(M * K, seed + 0);
-    const auto ref_rhs = fill_random<float>(N * K, seed + 1);
+
+    std::uniform_real_distribution<float> dist(-10.0, 1.0);
+    std::mt19937 rnd(seed + 1);
+    const auto ref_rhs = fill_matrix_raw<float>(1, N * K, [&dist, &rnd](size_t, size_t) { return dist(rnd); });
+
     const auto ref_biases = fill_random<float>(N, seed + 2);
 
     // Runs the reference implementation.
@@ -548,7 +551,11 @@ TEST_P(MatMulTest_f32_qai8dxp_qsi4cxp, EndToEnd_RHS_kxn_qsi4cx) {
 
     // Generates input data.
     const auto ref_lhs = fill_random<float>(M * K, seed + 0);
-    const auto ref_rhs = fill_random<float>(N * K, seed + 1);
+
+    std::uniform_real_distribution<float> dist(-10.0, 1.0);
+    std::mt19937 rnd(seed + 1);
+    const auto ref_rhs = fill_matrix_raw<float>(1, N * K, [&dist, &rnd](size_t, size_t) { return dist(rnd); });
+
     const auto ref_biases = fill_random<float>(N, seed + 2);
 
     // Transposed(nxk) RHS dimensions
@@ -679,7 +686,11 @@ TEST_P(MatMulTest_f32_qai8dxp_qsi4cxp, EndToEnd_RHS_kxn_qsu4cx) {
 
     // Generates input data.
     const auto ref_lhs = fill_random<float>(M * K, seed + 0);
-    const auto ref_rhs = fill_random<float>(N * K, seed + 1);
+
+    std::uniform_real_distribution<float> dist(-10.0, 1.0);
+    std::mt19937 rnd(seed + 1);
+    const auto ref_rhs = fill_matrix_raw<float>(1, N * K, [&dist, &rnd](size_t, size_t) { return dist(rnd); });
+
     const auto ref_biases = fill_random<float>(N, seed + 2);
 
     // Transposed(nxk) RHS dimensions