diff --git a/.ci/docker/common/install_onnx.sh b/.ci/docker/common/install_onnx.sh
index bc844ea1d40..f9502dc3550 100644
--- a/.ci/docker/common/install_onnx.sh
+++ b/.ci/docker/common/install_onnx.sh
@@ -16,7 +16,7 @@ pip_install \
   onnx==1.14.0
 
 pip_install \
-  onnxruntime==1.14.0 \
+  onnxruntime==1.15.0 \
   parameterized==0.8.1 \
   pytest-cov==4.0.0 \
   pytest-subtests==0.10.0 \
diff --git a/test/onnx/test_fx_op_consistency.py b/test/onnx/test_fx_op_consistency.py
index 828c22dc530..70f00ecde79 100644
--- a/test/onnx/test_fx_op_consistency.py
+++ b/test/onnx/test_fx_op_consistency.py
@@ -406,21 +406,6 @@ EXPECTED_SKIPS_OR_FAILS: Tuple[onnx_test_common.DecorateMeta, ...] = (
         dtypes=(torch.uint8, torch.int8, torch.int16,),
         reason=onnx_test_common.reason_onnx_script_does_not_support("Add", "int8, int16"),
     ),
-    xfail(
-        "new_full",
-        dtypes=onnx_test_common.BOOL_TYPES + onnx_test_common.INT_TYPES + onnx_test_common.FLOAT_TYPES,
-        reason=onnx_test_common.reason_onnx_runtime_does_not_support("new_full", "ORT 1.15 will support"),
-    ),
-    xfail(
-        "new_ones",
-        dtypes=onnx_test_common.BOOL_TYPES + onnx_test_common.INT_TYPES + onnx_test_common.FLOAT_TYPES,
-        reason=onnx_test_common.reason_onnx_runtime_does_not_support("new_ones", "ORT 1.15 will support"),
-    ),
-    xfail(
-        "new_zeros",
-        dtypes=onnx_test_common.BOOL_TYPES + onnx_test_common.INT_TYPES + onnx_test_common.FLOAT_TYPES,
-        reason=onnx_test_common.reason_onnx_runtime_does_not_support("new_zeros", "ORT 1.15 will support"),
-    ),
     xfail(
         "nn.functional.adaptive_avg_pool1d",
         reason=onnx_test_common.reason_onnx_script_does_not_support("aten.mean.dim"),
diff --git a/test/onnx/test_op_consistency.py b/test/onnx/test_op_consistency.py
index 8a0bc79e485..e90a00ef6ee 100644
--- a/test/onnx/test_op_consistency.py
+++ b/test/onnx/test_op_consistency.py
@@ -100,6 +100,8 @@ EXPECTED_SKIPS_OR_FAILS: Tuple[onnx_test_common.DecorateMeta, ...] = (
     skip("nn.functional.scaled_dot_product_attention", reason="fixme: ORT crashes on Windows, segfaults randomly on Linux"),
     skip("sqrt", dtypes=onnx_test_common.BOOL_TYPES, reason=onnx_test_common.reason_onnx_does_not_support("Sqrt")),
     skip("stft", opsets=[onnx_test_common.opsets_before(17)], reason=onnx_test_common.reason_onnx_does_not_support("STFT")),
+    xfail("stft",
+          reason=onnx_test_common.reason_onnx_runtime_does_not_support("STFT", "Regression on ORT=1.15 4 percent difference")),
     skip("tile", opsets=[onnx_test_common.opsets_before(13)], reason=onnx_test_common.reason_onnx_does_not_support("Tile")),
     xfail("unflatten", opsets=[onnx_test_common.opsets_before(13)], reason="Helper function is needed to support legacy ops."),
 )
diff --git a/test/onnx/test_pytorch_onnx_onnxruntime.py b/test/onnx/test_pytorch_onnx_onnxruntime.py
index bd51519c79d..b774822f233 100644
--- a/test/onnx/test_pytorch_onnx_onnxruntime.py
+++ b/test/onnx/test_pytorch_onnx_onnxruntime.py
@@ -650,15 +650,7 @@ class TestONNXRuntime(onnx_test_common._TestONNXRuntime):
 
         self.run_test(model, (x, y, z), input_names=("x", "y", "z"))
         self.run_test(model, (x,), {"y": y, "z": z}, input_names=("x", "y", "z"))
-
-        # Requires input_names to be set so that we can feed the inputs properly into ORT.
-        # TODO: Export default values as ONNX initializers, then this should not raise.
-        # https://msdata.visualstudio.com/Vienna/_workitems/edit/969268
-        # Default values are accessible via FunctionSchema.
-        with self.assertRaisesRegex(
-            ValueError, "Model requires 3 inputs. Input Feed contains 2"
-        ):
-            self.run_test(model, (x,), {"y": y}, input_names=("x", "y"))
+        self.run_test(model, (x,), {"y": y}, input_names=("x", "y"))
 
         for example_inputs, example_kwargs in (
             ((x, y, None), {}),
@@ -742,14 +734,13 @@ class TestONNXRuntime(onnx_test_common._TestONNXRuntime):
         y = torch.randn(2, 3)
         model = torch.jit.script(Model())
 
-        # TODO: Export default values as ONNX initializers, then this should not raise.
-        # https://msdata.visualstudio.com/Vienna/_workitems/edit/969268
-        # Default values are accessible via FunctionSchema.
-        with self.assertRaisesRegex(
-            ValueError, "Model requires 2 inputs. Input Feed contains 1"
-        ):
-            self.run_test(model, (x,))
-            self.run_test(model, (), {"y": y})
+        # Optional supports None inputs
+        self.run_test(model, (x,))
+        # NOTE: default value is not supported on ONNX, so torch and ONNX has
+        # different behavior
+        with self.assertRaisesRegex(AssertionError, "Tensor-likes are not close!"):
+            self.run_test(model, (), {"y": y}, input_names=["y"])
+
         self.run_test(model, (x, y))
         self.run_test(model, (), {"x": x, "y": y}, input_names=("x", "y"))