diff --git a/qlib/rl/trainer/vessel.py b/qlib/rl/trainer/vessel.py
index 6cd2eb3e9..b7912b488 100644
--- a/qlib/rl/trainer/vessel.py
+++ b/qlib/rl/trainer/vessel.py
@@ -168,7 +168,9 @@ class TrainingVessel(TrainingVesselBase):
         self.policy.train()
 
         with vector_env.collector_guard():
-            collector = Collector(self.policy, vector_env, VectorReplayBuffer(self.buffer_size, len(vector_env)))
+            collector = Collector(
+                self.policy, vector_env, VectorReplayBuffer(self.buffer_size, len(vector_env)), exploration_noise=True
+            )
 
             # Number of episodes collected in each training iteration can be overridden by fast dev run.
             if self.trainer.fast_dev_run is not None: