diff --git a/onnxscript/function_libs/torch_lib/ops/core.py b/onnxscript/function_libs/torch_lib/ops/core.py
index e7f1a611f7..87e4fc28f0 100644
--- a/onnxscript/function_libs/torch_lib/ops/core.py
+++ b/onnxscript/function_libs/torch_lib/ops/core.py
@@ -8062,10 +8062,46 @@ def aten_positive(self: TensorType) -> TensorType:
 @torch_op(("aten::pow.Tensor_Tensor", "_operator::pow"), trace_only=True)
 def aten_pow(self: TReal, exponent: TTensor) -> TReal:
     """pow(Tensor self, Tensor exponent) -> Tensor"""
-    # TODO(justinchuby): Add type promotion
+    if (
+        isinstance(self, ir.Value)
+        and isinstance(exponent, ir.Value)
+        and self.dtype != exponent.dtype
+        and (self.dtype.is_floating_point() or exponent.dtype.is_floating_point())
+    ):
+        # ONNX Pow keeps the dtype of the base, but torch promotes the result when
+        # the dtypes differ, e.g. an int64 base and a float32 exponent gives float32.
+        result_dtype = _promoted_float_dtype(self, exponent)
+        if self.dtype != result_dtype:
+            self = op.Cast(self, to=result_dtype)
+        # Pow accepts an integer exponent next to a float base, so integer exponents
+        # are left alone. Float and bool exponents (Pow has no bool input) follow the
+        # promoted dtype.
+        if not exponent.dtype.is_integer() and exponent.dtype != result_dtype:
+            exponent = op.Cast(exponent, to=result_dtype)
     return op.Pow(self, exponent)
 
 
+def _promoted_float_dtype(a: TensorType, b: TensorType) -> int:
+    """The result dtype torch gives to a binary op on tensors a and b, one of them float.
+
+    A zero-dim tensor only decides the dtype when it is of a higher category (float over
+    integer) than the other operand, mirroring torch.result_type.
+    """
+    a_float = a.dtype.is_floating_point()
+    b_float = b.dtype.is_floating_point()
+    if a_float != b_float:
+        # One float operand and one integer or bool operand: the float one decides.
+        return (a if a_float else b).dtype
+    a_scalar = a.shape is not None and len(a.shape) == 0
+    b_scalar = b.shape is not None and len(b.shape) == 0
+    if a_scalar != b_scalar:
+        # Both are floats of different widths. The dimensioned operand wins.
+        return (b if a_scalar else a).dtype
+    if {a.dtype, b.dtype} == {ir.DataType.FLOAT16, ir.DataType.BFLOAT16}:
+        return ir.DataType.FLOAT
+    return a.dtype if a.dtype.bitwidth >= b.dtype.bitwidth else b.dtype
+
+
 @torch_op("aten::pow.Tensor_Scalar", trace_only=True)
 def aten_pow_tensor_scalar(self: TReal, exponent: float) -> TReal:
     """pow(Tensor self, Scalar exponent) -> Tensor"""
diff --git a/tests/function_libs/torch_lib/e2e_ops_tests.py b/tests/function_libs/torch_lib/e2e_ops_tests.py
index 96cd89cf18..47b2040aca 100644
--- a/tests/function_libs/torch_lib/e2e_ops_tests.py
+++ b/tests/function_libs/torch_lib/e2e_ops_tests.py
@@ -45,6 +45,35 @@ def forward(self, x, indices, updates):
         onnx_program = torch.onnx.export(model, xs, dynamo=True)
         _testing.assert_onnx_program(onnx_program)
 
+    @parameterized.parameterized.expand(
+        [
+            ("int64_float32", torch.int64, torch.float32, (1, 3), (2, 1)),
+            ("float32_int64", torch.float32, torch.int64, (2, 1), (1, 3)),
+            ("float32_bool", torch.float32, torch.bool, (2, 1), (1, 3)),
+            ("float16_int64", torch.float16, torch.int64, (1, 3), (2, 1)),
+            ("float32_float64", torch.float32, torch.float64, (1, 3), (2, 1)),
+            ("float64_zero_dim_float32", torch.float64, torch.float32, (), (1, 3)),
+            ("int64_zero_dim_float64", torch.int64, torch.float64, (1, 3), ()),
+        ]
+    )
+    def test_pow_tensor_tensor_mixed_dtypes(
+        self, _: str, base_dtype, exponent_dtype, base_shape, exponent_shape
+    ):
+        class PowModel(torch.nn.Module):
+            def forward(self, x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
+                return torch.pow(x, y)
+
+        def make(dtype, shape):
+            return (
+                torch.randint(1, 4, shape)
+                if not dtype.is_floating_point
+                else torch.rand(shape) + 0.5
+            ).to(dtype)
+
+        inputs = (make(base_dtype, base_shape), make(exponent_dtype, exponent_shape))
+        onnx_program = torch.onnx.export(PowModel(), inputs, dynamo=True, optimize=False)
+        _testing.assert_onnx_program(onnx_program)
+
     def test_pow_tensor_scalar_int_float(self):
         class PowModel(torch.nn.Module):
             def forward(self, x: torch.Tensor) -> torch.Tensor: