diff --git a/paddle/framework/scope.h b/paddle/framework/scope.h
index 2ba3f8ed355b48800cfa4180e4e8a94f2c9958a9..c93b03e48130afe9568089b6a7586c4185d1d5b4 100644
--- a/paddle/framework/scope.h
+++ b/paddle/framework/scope.h
@@ -58,6 +58,8 @@ class Scope {
   /// nullptr if cannot find.
   Variable* FindVar(const std::string& name) const;
 
+  const Scope& parent() const { return *parent_; }
+
   /// Find the scope or an ancestor scope that contains the given variable.
   const Scope* FindScope(const Variable* var) const;
 
diff --git a/paddle/operators/prelu_op.h b/paddle/operators/prelu_op.h
index 63031c25cc3570cf40440726ea76976953d5417a..9843c476e4a49aba456219c0e2e992aa826f1685 100644
--- a/paddle/operators/prelu_op.h
+++ b/paddle/operators/prelu_op.h
@@ -94,7 +94,7 @@ class PReluGradKernel : public framework::OpKernel {
     Transform(context.device_context(), out_ptr, out_ptr + numel, dout_ptr,
               dx_ptr, PReluGradFunctor<T>(alpha_ptr));
 
-    // TODO (Zhuoyuan): add dalpha upgrade when GPU kernels ready
+    // TODO(Zhuoyuan): add dalpha upgrade when GPU kernels ready
   }
 };
 
diff --git a/paddle/operators/recurrent_op.cc b/paddle/operators/recurrent_op.cc
index fc57f0dcea9a5f81ada518bd5c945e9f50da5d7f..ad985839f5908d9235a4dbefc9b841362810114e 100644
--- a/paddle/operators/recurrent_op.cc
+++ b/paddle/operators/recurrent_op.cc
@@ -29,8 +29,11 @@ using Tensor = framework::Tensor;
 using LoDTensor = framework::LoDTensor;
 
 void RecurrentAlgorithm::InferShape(const Scope& scope) const {
-  seq_len_ =
-      scope.FindVar(arg_->inlinks[0])->GetMutable<LoDTensor>()->dims()[0];
+  auto* input0 = scope.FindVar(arg_->inlinks[0]);
+  PADDLE_ENFORCE_NOT_NULL(input0);
+  seq_len_ = input0->GetMutable<LoDTensor>()->dims()[0];
+  PADDLE_ENFORCE_GT(seq_len_, 0);
+
   CreateScopes(scope);
   auto step_scopes = GetStepScopes(scope);
   rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_,
diff --git a/paddle/operators/rnn/recurrent_op_utils.cc b/paddle/operators/rnn/recurrent_op_utils.cc
index 3e7e301ae3daa97fe0b802c44bfbd67620f7ddcb..ca7219b26d83eb6b8db75a5ed9cd360c5ac1d5df 100644
--- a/paddle/operators/rnn/recurrent_op_utils.cc
+++ b/paddle/operators/rnn/recurrent_op_utils.cc
@@ -28,14 +28,15 @@ void SegmentInputs(const std::vector<Scope*>& step_scopes,
                    const size_t seq_len, bool infer_shape_mode) {
   PADDLE_ENFORCE(!inlinks.empty(), "no in links are provided.");
   for (size_t i = 0; i < inlinks.size(); ++i) {
-    auto input_var = step_scopes[0]->FindVar(inlinks[i]);
-    PADDLE_ENFORCE(input_var != nullptr, "input link [%s] is not in scope.",
-                   inlinks[i]);
+    // global inputs
+    auto input_var = step_scopes[0]->parent().FindVar(inlinks[i]);
+    PADDLE_ENFORCE_NOT_NULL(input_var, "input link [%s] is not in scope.",
+                            inlinks[i]);
 
     LoDTensor* input = input_var->GetMutable<LoDTensor>();
     f::DDim dims = input->dims();
-    PADDLE_ENFORCE(static_cast<size_t>(dims[0]) == seq_len,
-                   "all the inlinks must have same length");
+    PADDLE_ENFORCE_EQ(static_cast<size_t>(dims[0]), seq_len,
+                      "all the inlinks be the same length");
     f::DDim step_dims = slice_ddim(dims, 1, dims.size());
     for (size_t j = 0; j < seq_len; j++) {
       Tensor* step_input =
@@ -54,15 +55,14 @@ void ConcatOutputs(const std::vector<Scope*>& step_scopes,
                    const std::vector<std::string>& outlinks,
                    const size_t seq_len, bool infer_shape_mode) {
   for (size_t i = 0; i < outlinks.size(); i++) {
-    auto output_var = step_scopes[0]->FindVar(outlinks[i]);
-    PADDLE_ENFORCE(output_var != nullptr, "output link [%s] is not in scope.",
-                   outlinks[i]);
+    auto output_var = step_scopes[0]->parent().FindVar(outlinks[i]);
+    PADDLE_ENFORCE_NOT_NULL(output_var, "output link [%s] is not in scope.",
+                            outlinks[i]);
     LoDTensor* output = output_var->GetMutable<LoDTensor>();
 
     if (infer_shape_mode) {
-      auto step_scope_var = step_scopes[0]->FindVar(outlinks[i].internal);
-      PADDLE_ENFORCE(step_scope_var != nullptr, "%s not in scope",
-                     outlinks[i].internal);
+      auto step_scope_var = step_scopes[0]->FindVar(outlinks[i]);
+      PADDLE_ENFORCE_NOT_NULL(step_scope_var, "%s not in scope", outlinks[i]);
       f::DDim step_dims =
           step_scope_var->template GetMutable<LoDTensor>()->dims();
       std::vector<int64_t> dims_vec = vectorize(step_dims);
diff --git a/python/paddle/v2/framework/tests/test_recurrent_op.py b/python/paddle/v2/framework/tests/test_recurrent_op.py
index 886f941d4dc5fa253e92a4a7ac0c5e5abf7b0797..79eda70021b76cd06e4c40740b1ca49476f4c503 100644
--- a/python/paddle/v2/framework/tests/test_recurrent_op.py
+++ b/python/paddle/v2/framework/tests/test_recurrent_op.py
@@ -59,7 +59,6 @@ class PySimpleRNNTest(unittest.TestCase):
 
     def test_forward(self):
         output = self.rnn.forward()
-        print 'output', output
 
 
 def create_tensor(scope, name, shape, np_data):
@@ -103,7 +102,7 @@ class TestRecurrentOp(unittest.TestCase):
         ctx = core.DeviceContext.create(core.CPUPlace())
         self.rnnop.infer_shape(self.scope)
         self.rnnop.run(self.scope, ctx)
-        return np.array(self.scope.find_var("h").get_tensor())
+        return np.array(self.scope.find_var("h@mem").get_tensor())
 
     def create_global_variables(self):
         # create inlink
@@ -123,7 +122,7 @@ class TestRecurrentOp(unittest.TestCase):
         create_tensor(self.scope, "h_boot", [self.batch_size, self.input_dim],
                       h_boot_np_data)
         self.scope.new_var("step_scopes")
-        self.scope.new_var("h")
+        self.scope.new_var("h@mem")
 
     def create_rnn_op(self):
         # create RNNOp
@@ -133,7 +132,7 @@ class TestRecurrentOp(unittest.TestCase):
             boot_memories=["h_boot"],
             step_net="stepnet",
             # outputs
-            outlinks=["h"],
+            outlinks=["h@mem"],
             step_scopes="step_scopes",
             # attributes
             pre_memories=["h@pre"],