diff --git a/swift/llm/argument/rlhf_args.py b/swift/llm/argument/rlhf_args.py
index 68dbd9f7d6..1c53e48ab8 100644
--- a/swift/llm/argument/rlhf_args.py
+++ b/swift/llm/argument/rlhf_args.py
@@ -43,6 +43,9 @@ class RLHFArguments(TrainArguments):
     desirable_weight: float = 1.0
     undesirable_weight: float = 1.0
 
+    # Use last_round by default
+    loss_scale: str = 'last_round'
+
     def __post_init__(self):
         self._init_simpo()
         self._set_default()
diff --git a/swift/llm/train/rlhf.py b/swift/llm/train/rlhf.py
index 906a3e1166..a11a32aed5 100644
--- a/swift/llm/train/rlhf.py
+++ b/swift/llm/train/rlhf.py
@@ -28,9 +28,10 @@ def _prepare_template(self) -> None:
         mode = 'kto' if args.rlhf_type == 'kto' else 'rlhf'
         self.template.set_mode(mode)
 
-        if args.rlhf_type != 'orpo' or args.model_meta.is_multimodal:
+        if args.rlhf_type == 'orpo' and not args.model_meta.is_multimodal:
             # Avoid padding labels during the model's forward pass in multimodal models.
-            self.template.loss_scale = 'last_round'
+            args.loss_scale = 'default'
+        self.template.loss_scale = args.loss_scale
 
     @classmethod
     def prepare_model(cls, args, model, *_args, **kwargs):