diff --git a/mllib/src/main/scala/org/apache/spark/ml/optim/aggregator/HuberAggregator.scala b/mllib/src/main/scala/org/apache/spark/ml/optim/aggregator/HuberAggregator.scala
index 13f64d2d50424..fc4c423a60b2a 100644
--- a/mllib/src/main/scala/org/apache/spark/ml/optim/aggregator/HuberAggregator.scala
+++ b/mllib/src/main/scala/org/apache/spark/ml/optim/aggregator/HuberAggregator.scala
@@ -81,6 +81,8 @@ private[ml] class HuberAggregator(
   } else {
     0.0
   }
+  // make transient so we do not serialize between aggregation stages
+  @transient private lazy val coefficients = bcParameters.value.toArray.slice(0, numFeatures)
 
   /**
    * Add a new training instance to this HuberAggregator, and update the loss and gradient
@@ -97,7 +99,7 @@ private[ml] class HuberAggregator(
 
       if (weight == 0.0) return this
       val localFeaturesStd = bcFeaturesStd.value
-      val localCoefficients = bcParameters.value.toArray.slice(0, numFeatures)
+      val localCoefficients = coefficients
       val localGradientSumArray = gradientSumArray
 
       val margin = {