diff --git a/onnxruntime/test/providers/cpu/ml/write_scores_test.cc b/onnxruntime/test/providers/cpu/ml/write_scores_test.cc
new file mode 100644
index 0000000000..932f460793
--- /dev/null
+++ b/onnxruntime/test/providers/cpu/ml/write_scores_test.cc
@@ -0,0 +1,87 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+#include "gtest/gtest.h"
+#include "core/framework/tensor.h"
+#include "core/providers/cpu/ml/ml_common.h"
+#include "test/framework/dummy_allocator.h"
+#include <random>
+
+using namespace onnxruntime;
+using namespace onnxruntime::ml;
+class WriteScores : public ::testing::Test {
+ protected:
+  std::random_device r;
+  std::default_random_engine rd{r()};
+};
+
+TEST_F(WriteScores, multiple_scores_transform_none) {
+  POST_EVAL_TRANSFORM trans[] = {POST_EVAL_TRANSFORM::NONE,
+                                 POST_EVAL_TRANSFORM::LOGISTIC,
+                                 POST_EVAL_TRANSFORM::SOFTMAX,
+                                 POST_EVAL_TRANSFORM::SOFTMAX_ZERO,
+                                 POST_EVAL_TRANSFORM::PROBIT};
+  for (POST_EVAL_TRANSFORM tran : trans) {
+    std::vector<float> v1(100);
+    std::uniform_real_distribution<float> uniform_dist(-1, 1);
+    std::generate_n(v1.data(), v1.size(), [&]() -> float { return uniform_dist(rd); });
+    std::vector<float> v2 = v1;
+    auto alloc = std::make_shared<test::DummyAllocator>();
+    Tensor t(DataTypeImpl::GetType<float>(), {static_cast<int64_t>(v1.size())}, alloc);
+    if (tran == POST_EVAL_TRANSFORM::SOFTMAX_ZERO) {
+      //random set one number as zero
+      v1[3] = 0;
+    }
+    write_scores<float>(v1, tran, 0, &t, -1);
+    const float* output_data = t.Data<float>();
+    //verify the result
+    if (tran == POST_EVAL_TRANSFORM::NONE) {
+      for (size_t i = 0; i != v2.size(); ++i) {
+        EXPECT_FLOAT_EQ(v1[i], output_data[i]);
+      }
+    } else if (tran == POST_EVAL_TRANSFORM::SOFTMAX || tran == POST_EVAL_TRANSFORM::SOFTMAX_ZERO) {
+      //sum of the output values should be near to 1
+      double sum = 0;
+      for (size_t i = 0; i != v1.size(); ++i) {
+        sum += output_data[i];
+      }
+      EXPECT_NEAR(sum, 1, 1e-6);
+
+      if (tran == POST_EVAL_TRANSFORM::SOFTMAX_ZERO) {
+        EXPECT_FLOAT_EQ(output_data[3], 0);
+      }
+    }
+  }
+}
+
+TEST_F(WriteScores, single_score_transform_none) {
+  std::vector<float> v1;
+  std::uniform_real_distribution<float> uniform_dist(-100, 100);
+  v1.push_back(uniform_dist(rd));
+  std::vector<float> v2 = v1;
+  auto alloc = std::make_shared<test::DummyAllocator>();
+  Tensor t(DataTypeImpl::GetType<float>(), {static_cast<int64_t>(v1.size())}, alloc);
+  write_scores<float>(v1, POST_EVAL_TRANSFORM::NONE, 0, &t, -1);
+  const float* output_data = t.Data<float>();
+  for (size_t i = 0; i != v2.size(); ++i) {
+    EXPECT_FLOAT_EQ(v1[i], output_data[i]);
+  }
+}
+
+TEST_F(WriteScores, single_score_transform_none_add_second_class) {
+  for (int i = 0; i != 4; ++i) {
+    std::vector<float> v1;
+    std::uniform_real_distribution<float> uniform_dist(-100, 100);
+    v1.push_back(uniform_dist(rd));
+    std::vector<float> v2 = v1;
+    auto alloc = std::make_shared<test::DummyAllocator>();
+    Tensor t(DataTypeImpl::GetType<float>(), {2}, alloc);
+    write_scores<float>(v1, POST_EVAL_TRANSFORM::NONE, 0, &t, i);
+    const float* output_data = t.Data<float>();
+	EXPECT_NEAR(output_data[1] + output_data[0], i == 0 || i == 1 ? 1 : 0, 1e-6);
+    if (i == 0 || i == 1) {
+      EXPECT_FLOAT_EQ(v2[0], output_data[1]);
+      EXPECT_FLOAT_EQ(1 - v2[0], output_data[0]);
+    }
+  }
+}
\ No newline at end of file