diff --git a/cmake/generic.cmake b/cmake/generic.cmake
index 83e3d155d038cc65f3e372f0e4ba0aaee2e29690..a30cdeff62562fb349e30e0dce073408ccbcc48d 100644
--- a/cmake/generic.cmake
+++ b/cmake/generic.cmake
@@ -93,7 +93,7 @@ include_directories(${CMAKE_CURRENT_BINARY_DIR})
 if(NOT APPLE)
     find_package(Threads REQUIRED)
     link_libraries(${CMAKE_THREAD_LIBS_INIT})
-    set(CMAKE_CXX_LINK_EXECUTABLE "${CMAKE_CXX_LINK_EXECUTABLE} -ldl")
+    set(CMAKE_CXX_LINK_EXECUTABLE "${CMAKE_CXX_LINK_EXECUTABLE} -ldl -lrt")
 endif(NOT APPLE)
 
 function(merge_static_libs TARGET_NAME)
diff --git a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp b/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp
index 4cb5b8ec2dbb93cc17075079754c2f1b8cd392e5..8cc08eb20857daaf98e40afe0dea8ba3a35f90e0 100644
--- a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp
+++ b/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp
@@ -1395,17 +1395,23 @@ void RecurrentGradientMachine::createDataOutlinkCopySizeInfo(
   batchMachineStartPos_.resize(totalSeqNum + 1, 0);
   if (isSeq) {
     ICpuGpuVectorPtr inputSeqStartPos = outArgs[0].sequenceStartPositions;
-    CHECK_EQ(inputSeqStartPos->getSize() - 1, finalPaths_.size());
+    CHECK_EQ(static_cast<size_t>(inputSeqStartPos->getSize() - 1),
+             getBeamSize() > 1 ? finalPaths_.size() : finalPaths_[0].size());
     int* starts = inputSeqStartPos->getMutableData(false);
     int seqId = 0;
     for (int i = 0; i < finalPaths_.size(); ++i) {
       for (int j = 0; j < finalPaths_[i].size(); ++j) {
-        copySize[seqId] = starts[i + 1] - starts[i];
+        copySize[seqId] = getBeamSize() > 1 ? starts[i + 1] - starts[i]
+                                            : starts[j + 1] - starts[j];
         batchMachineStartPos_[seqId + 1] =
             batchMachineStartPos_[seqId] + finalPaths_[i][j].ids.size();
         seqId++;
       }
     }
+  } else {
+    for (size_t i = 0; i < finalPaths_[0].size(); ++i)
+      batchMachineStartPos_[i + 1] =
+          batchMachineStartPos_[i] + finalPaths_[0][i].ids.size();
   }
 }
 
diff --git a/paddle/gserver/gradientmachines/RecurrentGradientMachine.h b/paddle/gserver/gradientmachines/RecurrentGradientMachine.h
index bd096770b75b4f44488777c0c0e27b14f86b2015..a3d04b207c0d84ea9c77c646314efedcd49772b1 100644
--- a/paddle/gserver/gradientmachines/RecurrentGradientMachine.h
+++ b/paddle/gserver/gradientmachines/RecurrentGradientMachine.h
@@ -477,8 +477,6 @@ private:
    * outlinks.
    * @note In beam search, only one generated sequence with the hightest log
    * probabilites are retained.
-   * @param machineIdVec : select a row of output matrix in each frame
-   * that the generation process expanded.
    */
   void createDataOutlink();
   void createDataOutlinkCopySizeInfo(bool isSeq,
diff --git a/paddle/parameter/Argument.cpp b/paddle/parameter/Argument.cpp
index e7522def08cb32b5585e676211719621883c70b8..f45a51d7b1836f02945f5efa3b061abc5fe7f974 100644
--- a/paddle/parameter/Argument.cpp
+++ b/paddle/parameter/Argument.cpp
@@ -310,8 +310,8 @@ void Argument::concat(const std::vector<Argument>& args,
 
   auto copyIds = [batchSize, stream](IVectorPtr& dst,
                                      const IVectorPtr& src,
-                                     int startRow,
-                                     int pos,
+                                     int desStartRow,
+                                     int srcStartRow,
                                      int size,
                                      bool useGpu) {
     if (!src) {
@@ -319,13 +319,14 @@ void Argument::concat(const std::vector<Argument>& args,
       return;
     }
     IVector::resizeOrCreate(dst, batchSize, useGpu);
-    dst->subVec(startRow, size)->copyFrom(*src->subVec(pos, size), stream);
+    dst->subVec(desStartRow, size)
+        ->copyFrom(*src->subVec(srcStartRow, size), stream);
   };
 
   auto copyStrs = [batchSize, stream](SVectorPtr& dst,
                                       const SVectorPtr& src,
-                                      int startRow,
-                                      int pos,
+                                      int desStartRow,
+                                      int srcStartRow,
                                       int size,
                                       bool useGpu) {
     if (!src) {
@@ -337,8 +338,9 @@ void Argument::concat(const std::vector<Argument>& args,
     } else {
       dst->resize(batchSize);
     }
-    std::copy(
-        src->begin() + pos, src->begin() + pos + size, dst->begin() + startRow);
+    std::copy(src->begin() + srcStartRow,
+              src->begin() + srcStartRow + size,
+              dst->begin() + desStartRow);
   };
 
   dataId = args[0].dataId;
diff --git a/python/paddle/trainer_config_helpers/networks.py b/python/paddle/trainer_config_helpers/networks.py
index 6cddf547a72337cd86e566bca93495f76a8716bc..30c826ffc8e7e5ef384520e6c95cd69584df971d 100755
--- a/python/paddle/trainer_config_helpers/networks.py
+++ b/python/paddle/trainer_config_helpers/networks.py
@@ -1370,7 +1370,14 @@ def simple_attention(encoded_sequence,
         param_attr=softmax_param_attr,
         name="%s_softmax" % name,
         bias_attr=False)
-    return attention_weight
+
+    scaled = scaling_layer(
+        weight=attention_weight,
+        input=encoded_sequence,
+        name='%s_scaling' % name)
+    return pooling_layer(
+        input=scaled, pooling_type=SumPooling(),
+        name="%s_pooling" % name), attention_weight
 
 
 def inputs(layers, *args):