diff --git a/sagemaker-train/tests/unit/train/test_model_trainer_pipeline_variable.py b/sagemaker-train/tests/unit/train/test_model_trainer_pipeline_variable.py
index 3fd34fa47b..24f7d2184e 100644
--- a/sagemaker-train/tests/unit/train/test_model_trainer_pipeline_variable.py
+++ b/sagemaker-train/tests/unit/train/test_model_trainer_pipeline_variable.py
@@ -26,7 +26,7 @@
 
 from sagemaker.core.helper.session_helper import Session
 from sagemaker.core.helper.pipeline_variable import PipelineVariable, StrPipeVar
-from sagemaker.core.workflow.parameters import ParameterString
+from sagemaker.core.workflow.parameters import ParameterString, ParameterInteger
 from sagemaker.train.model_trainer import ModelTrainer, Mode
 from sagemaker.train.configs import (
     Compute,
@@ -116,6 +116,123 @@ def test_environment_values_accept_parameter_string(self):
         assert trainer.environment["DATASET_VERSION"] is param
         assert trainer.environment["STATIC_VAR"] == "hello"
 
+    def test_hyperparameters_accept_parameter_integer(self):
+        """ModelTrainer.hyperparameters should accept ParameterInteger values (GH#5504)."""
+        param = ParameterInteger(name="MaxDepth", default_value=5)
+        trainer = ModelTrainer(
+            training_image=DEFAULT_IMAGE,
+            hyperparameters={"max_depth": param},
+            base_job_name="pipeline-test-job",
+            role=DEFAULT_ROLE,
+            compute=DEFAULT_COMPUTE,
+            stopping_condition=DEFAULT_STOPPING,
+            output_data_config=DEFAULT_OUTPUT,
+        )
+        assert trainer.hyperparameters["max_depth"] is param
+
+    def test_hyperparameters_accept_parameter_string(self):
+        """ModelTrainer.hyperparameters should accept ParameterString values (GH#5504)."""
+        param = ParameterString(name="Algorithm", default_value="xgboost")
+        trainer = ModelTrainer(
+            training_image=DEFAULT_IMAGE,
+            hyperparameters={"algorithm": param},
+            base_job_name="pipeline-test-job",
+            role=DEFAULT_ROLE,
+            compute=DEFAULT_COMPUTE,
+            stopping_condition=DEFAULT_STOPPING,
+            output_data_config=DEFAULT_OUTPUT,
+        )
+        assert trainer.hyperparameters["algorithm"] is param
+
+    def test_hyperparameters_accept_mixed_pipeline_and_plain_values(self):
+        """ModelTrainer.hyperparameters should accept a mix of PipelineVariable and plain values.
+
+        Regression test for GH#5504.
+        """
+        param_int = ParameterInteger(name="MaxDepth", default_value=5)
+        param_str = ParameterString(name="Objective", default_value="reg:squarederror")
+        trainer = ModelTrainer(
+            training_image=DEFAULT_IMAGE,
+            hyperparameters={
+                "max_depth": param_int,
+                "objective": param_str,
+                "eta": 0.1,
+                "num_round": "100",
+            },
+            base_job_name="pipeline-test-job",
+            role=DEFAULT_ROLE,
+            compute=DEFAULT_COMPUTE,
+            stopping_condition=DEFAULT_STOPPING,
+            output_data_config=DEFAULT_OUTPUT,
+        )
+        assert trainer.hyperparameters["max_depth"] is param_int
+        assert trainer.hyperparameters["objective"] is param_str
+        assert trainer.hyperparameters["eta"] == 0.1
+        assert trainer.hyperparameters["num_round"] == "100"
+
+
+class TestSafeSerializePipelineVariable:
+    """Test that safe_serialize correctly preserves PipelineVariable objects (GH#5504)."""
+
+    def test_safe_serialize_preserves_parameter_integer(self):
+        """safe_serialize should return PipelineVariable as-is, not stringify it."""
+        from sagemaker.train.utils import safe_serialize
+
+        param = ParameterInteger(name="MaxDepth", default_value=5)
+        result = safe_serialize(param)
+        assert result is param
+        assert isinstance(result, PipelineVariable)
+
+    def test_safe_serialize_preserves_parameter_string(self):
+        """safe_serialize should return ParameterString as-is."""
+        from sagemaker.train.utils import safe_serialize
+
+        param = ParameterString(name="Objective", default_value="reg:squarederror")
+        result = safe_serialize(param)
+        assert result is param
+        assert isinstance(result, PipelineVariable)
+
+    def test_safe_serialize_still_serializes_plain_values(self):
+        """safe_serialize should still JSON-serialize plain values."""
+        from sagemaker.train.utils import safe_serialize
+
+        assert safe_serialize(42) == "42"
+        assert safe_serialize("hello") == '"hello"'
+        assert safe_serialize(0.1) == "0.1"
+
+    def test_create_training_job_args_preserves_pipeline_hyperparameters(
+        self,
+    ):
+        """_create_training_job_args should preserve PipelineVariable in hyperparameters.
+
+        Regression test for GH#5504.
+        """
+        param_int = ParameterInteger(name="MaxDepth", default_value=5)
+        param_str = ParameterString(
+            name="Objective", default_value="reg:squarederror"
+        )
+        trainer = ModelTrainer(
+            training_image=DEFAULT_IMAGE,
+            hyperparameters={
+                "max_depth": param_int,
+                "objective": param_str,
+                "eta": 0.1,
+                "num_round": "100",
+            },
+            base_job_name="pipeline-test-job",
+            role=DEFAULT_ROLE,
+            compute=DEFAULT_COMPUTE,
+            stopping_condition=DEFAULT_STOPPING,
+            output_data_config=DEFAULT_OUTPUT,
+        )
+        args = trainer._create_training_job_args()
+        hp = args["hyperparameters"]
+        assert hp["max_depth"] is param_int
+        assert hp["objective"] is param_str
+        # Plain values should be JSON-serialized strings
+        assert hp["eta"] == "0.1"
+        assert hp["num_round"] == '"100"'
+
 
 class TestModelTrainerRealValuesStillWork:
     """Regression tests: verify that passing real values still works after the change."""