diff --git a/src/operator/softmax_output-inl.h b/src/operator/softmax_output-inl.h
index 5f8203e824a3..e19dbe259dff 100644
--- a/src/operator/softmax_output-inl.h
+++ b/src/operator/softmax_output-inl.h
@@ -53,6 +53,7 @@ struct SoftmaxOutputParam : public dmlc::Parameter<SoftmaxOutputParam> {
   bool preserve_shape;
   int normalization;
   bool out_grad;
+  float smooth_alpha;
   DMLC_DECLARE_PARAMETER(SoftmaxOutputParam) {
     DMLC_DECLARE_FIELD(grad_scale).set_default(1.0f)
     .describe("Scales the gradient by a float factor.");
@@ -78,6 +79,10 @@ struct SoftmaxOutputParam : public dmlc::Parameter<SoftmaxOutputParam> {
     DMLC_DECLARE_FIELD(out_grad)
     .set_default(false)
     .describe("Multiplies gradient with output gradient element-wise.");
+    DMLC_DECLARE_FIELD(smooth_alpha)
+    .set_default(0.0f)
+    .set_range(0.0f,1.0f)
+    .describe("Constant for smoothed cross-entropy gradients.");
   };
 };
 
@@ -215,9 +220,17 @@ class SoftmaxOutputOp : public Operator {
           in_grad[softmaxout_enum::kData].get_with_shape<xpu, 2, DType>(data_shape, s);
       index_t valid_cnt = label.shape_.Size();
       if (param_.use_ignore) {
-        SoftmaxGrad(grad, out, label, static_cast<DType>(param_.ignore_label));
+        if (param_.smooth_alpha == 0.0f) {
+          SoftmaxGrad(grad, out, label, static_cast<DType>(param_.ignore_label));
+        } else {
+          SmoothSoftmaxGrad(grad, out, label, static_cast<DType>(param_.ignore_label), param_.smooth_alpha);
+        }
       } else {
-        SoftmaxGrad(grad, out, label);
+        if (param_.smooth_alpha == 0.0f) {
+          SoftmaxGrad(grad, out, label);
+        } else {
+          SmoothSoftmaxGrad(grad, out, label, param_.smooth_alpha);
+        }
       }
       if (param_.normalization == softmaxout_enum::kBatch) {
         valid_cnt = label.size(0);
diff --git a/tests/python/unittest/test_operator.py b/tests/python/unittest/test_operator.py
index 0c528e0b180f..fd88bb42ea75 100644
--- a/tests/python/unittest/test_operator.py
+++ b/tests/python/unittest/test_operator.py
@@ -235,6 +235,55 @@ def test_regression():
                      lambda x, y : x - y)
 
 
+def check_softmax_grad():
+    x = mx.sym.Variable('x')
+    label = mx.sym.Variable('label')
+    x_nd = mx.nd.array([[1, 6, 4, 2]], ctx=default_context())
+    grad_x = mx.nd.zeros((1,4), ctx=default_context())
+    label_nd = mx.nd.array([1], ctx=default_context())
+
+    sym = mx.sym.SoftmaxOutput(data=x, label=label, ignore_label=0, use_ignore=False)
+    ex = sym.bind(ctx=default_context(), args={'x': x_nd, 'label': label_nd}, args_grad={'x': grad_x})
+
+    ex.forward(is_train=True)
+    softmax_out = ex.outputs[0].asnumpy()
+    expected_softmax_out = mx.nd.softmax(x_nd).asnumpy()
+    print(softmax_out)
+    assert np.isclose(softmax_out, expected_softmax_out).all()
+
+    ex.backward(is_train=True)
+    grad_out = ex.grad_arrays[0].asnumpy()
+    k = int(label_nd[0].asscalar())
+    expected_grad_out = np.zeros((1,4))
+    expected_grad_out[0, k] = -1
+    assert np.isclose(grad_out - softmax_out, expected_grad_out).all()
+
+
+def check_smoothed_softmax_grad():
+    alpha = 0.2
+    x = mx.sym.Variable('x')
+    label = mx.sym.Variable('label')
+    x_nd = mx.nd.array([[1, 6, 4, 2]], ctx=default_context())
+    grad_x = mx.nd.zeros((1,4), ctx=default_context())
+    label_nd = mx.nd.array([1], ctx=default_context())
+
+    sym = mx.sym.SoftmaxOutput(data=x, label=label, ignore_label=0, use_ignore=False, smooth_alpha=alpha)
+    ex = sym.bind(ctx=default_context(), args={'x': x_nd, 'label': label_nd}, args_grad={'x': grad_x})
+
+    ex.forward(is_train=True)
+    softmax_out = ex.outputs[0].asnumpy()
+    expected_softmax_out = mx.nd.softmax(x_nd).asnumpy()
+    print(softmax_out)
+    assert np.isclose(softmax_out, expected_softmax_out).all()
+
+    ex.backward(is_train=True)
+    grad_out = ex.grad_arrays[0].asnumpy()
+    k = int(label_nd[0].asscalar())
+    expected_grad_out = np.full((1,4), fill_value=-alpha/(4-1))
+    expected_grad_out[0, k] = - (1 - alpha)
+    assert np.isclose(grad_out - softmax_out, expected_grad_out).all()
+
+
 def check_softmax_with_ignore_label(xpu):
     X = mx.symbol.Variable('X')
     L = mx.symbol.Variable('L')
@@ -289,6 +338,8 @@ def test_softmax():
     check_softmax_with_shape((3, 4), default_context(), preserve_shape=False)
     check_softmax_with_shape((3, 4), default_context(), preserve_shape=True)
     check_softmax_with_shape((3, 4, 2), default_context(), preserve_shape=True)
+    check_smoothed_softmax_grad()
+    check_smoothed_softmax_grad()
 
 
 def test_python_op():