diff --git a/flaml/tune/searcher/flow2.py b/flaml/tune/searcher/flow2.py
index 7ccf6fa6bb..371b503273 100644
--- a/flaml/tune/searcher/flow2.py
+++ b/flaml/tune/searcher/flow2.py
@@ -674,5 +674,8 @@ def reach(self, other: Searcher) -> bool:
             # unordered cat choice is hard to reach by chance
             if config1[key] != config2.get(key):
                 return False
-        delta = np.array([incumbent1[key] - incumbent2.get(key, np.inf) for key in self._tunable_keys])
+        try:
+            delta = np.array([incumbent1[key] - incumbent2.get(key, np.inf) for key in self._tunable_keys])
+        except TypeError:
+            return False
         return np.linalg.norm(delta) <= self.step
diff --git a/test/tune/test_searcher.py b/test/tune/test_searcher.py
index e49c1aa5c1..de33e545cf 100644
--- a/test/tune/test_searcher.py
+++ b/test/tune/test_searcher.py
@@ -345,5 +345,27 @@ def test_unresolved_search_space(caplog):
     ), "BlendSearch should not produce warning about unresolved search space"
 
 
+def test_flow2_reach_mixed_type_incumbents():
+    """Regression test for #903.
+
+    FLOW2.reach() must return False instead of raising TypeError when conditional choice parameters lead to mixed-type incumbents.
+    """
+    from flaml.tune.searcher.flow2 import FLOW2
+
+    f1 = FLOW2.__new__(FLOW2)
+    f1.best_config = {"x": 1.0}
+    f1.incumbent = {"x": 1.0}
+    f1._resource = None
+    f1._unordered_cat_hp = {}
+    f1._tunable_keys = ["x"]
+    f1.step = 0.5
+
+    f2 = FLOW2.__new__(FLOW2)
+    f2.best_config = {"x": "None"}
+    f2.incumbent = {"x": "None"}
+
+    assert f1.reach(f2) is False
+
+
 if __name__ == "__main__":
     test_unresolved_search_space(None)