diff --git a/onnxscript/rewriter/rules/common/_basic_rules_test.py b/onnxscript/rewriter/rules/common/_basic_rules_test.py
index 7d4e9d9b33..67ebdaa495 100644
--- a/onnxscript/rewriter/rules/common/_basic_rules_test.py
+++ b/onnxscript/rewriter/rules/common/_basic_rules_test.py
@@ -528,7 +528,11 @@ def test_reshape_reshape_dynamic_rule(self, input_shape, shape2, allowzero2=0):
 
         # Check inference.
         inputs = np.random.default_rng(7).random(input_shape, dtype="float32")
-        testing.assert_numerically_equal(model, updated_model, (inputs,), atol=0, rtol=0)
+        # Use the reference implementation to avoid ORT incorrectly folding/rewriting
+        # the original two-reshape model (e.g. ignoring allowzero=1).
+        testing.assert_numerically_equal(
+            model, updated_model, (inputs,), atol=0, rtol=0, use_reference=True
+        )
 
     @parameterized.parameterized.expand(
         [
diff --git a/tests/function_libs/torch_lib/ops_test_data.py b/tests/function_libs/torch_lib/ops_test_data.py
index 4e1fcae128..4bf98f587d 100644
--- a/tests/function_libs/torch_lib/ops_test_data.py
+++ b/tests/function_libs/torch_lib/ops_test_data.py
@@ -838,7 +838,9 @@ def _where_input_wrangler(
         "logcumsumexp", core_ops.aten_logcumsumexp, tolerance={torch.float16: (1e-2, 1e-1)}
     ),
     TorchLibOpInfo("logdet", core_ops.aten_logdet),
-    TorchLibOpInfo("logsumexp", core_ops.aten_logsumexp),
+    TorchLibOpInfo(
+        "logsumexp", core_ops.aten_logsumexp, tolerance={torch.float16: (2e-2, 1e-4)}
+    ),
     TorchLibOpInfo("lt", core_ops.aten_lt),
     TorchLibOpInfo("masked_fill", core_ops.aten_masked_fill).xfail(
         dtypes=(torch.bool,),