diff --git a/paimon-python/pypaimon/multimodal/window_dataset.py b/paimon-python/pypaimon/multimodal/window_dataset.py
index 4a8209c955ee..51ffe7ecf715 100644
--- a/paimon-python/pypaimon/multimodal/window_dataset.py
+++ b/paimon-python/pypaimon/multimodal/window_dataset.py
@@ -46,6 +46,19 @@
 from pypaimon.utils.range import Range
 
 
+_IMMUTABLE_CELL_TYPES = frozenset((bool, int, float, str, bytes, type(None)))
+
+
+def _copy_window_value(value):
+    # Arrow numeric vectors are flat lists. Copy the container without walking
+    # deepcopy's memo machinery for every scalar in overlapping windows.
+    if type(value) in _IMMUTABLE_CELL_TYPES:
+        return value
+    if type(value) is list and all(type(item) in _IMMUTABLE_CELL_TYPES for item in value):
+        return value.copy()
+    return copy.deepcopy(value)
+
+
 class ContiguousWindowDataset(Dataset):
     """Map-style training samples from a single table of contiguous frames.
 
@@ -310,7 +323,7 @@ def _sample(self, anchor, position, reads):
         for name in self.columns:
             plans, rows = reads[name]
             row_ids, padding = plans[position]
-            values = [copy.deepcopy(
+            values = [_copy_window_value(
                 self.pad_values[name] if row_id is None else rows[row_id][name])
                 for row_id in row_ids]
             if self.window_size is None:
diff --git a/paimon-python/pypaimon/tests/contiguous_window_dataset_test.py b/paimon-python/pypaimon/tests/contiguous_window_dataset_test.py
index ac3296295d83..5aad9ab7fd1c 100644
--- a/paimon-python/pypaimon/tests/contiguous_window_dataset_test.py
+++ b/paimon-python/pypaimon/tests/contiguous_window_dataset_test.py
@@ -614,11 +614,13 @@ def test_plural_access_isolates_mutable_cells_between_samples(self):
                 pa.field("episode", pa.string(), nullable=False),
                 pa.field("step", pa.int32(), nullable=False),
                 pa.field("values", pa.list_(pa.int32()), nullable=False),
+                pa.field("nested", pa.list_(pa.list_(pa.int32())), nullable=False),
             ]),
             options=_TABLE_OPTIONS,
         )
         table.add([
-            {"episode": "episode-a", "step": step, "values": [step]}
+            {"episode": "episode-a", "step": step,
+             "values": [step], "nested": [[step]]}
             for step in range(3)
         ])
 
@@ -627,21 +629,36 @@ def mutate(values):
                 value.append(99)
             return values
 
+        def mutate_nested(values):
+            for value in values:
+                value[0].append(99)
+            return values
+
+        padding = {"values": [-1], "nested": [[-1]]}
         dataset = table.scan().to_contiguous_window_dataset(
-            window_size=2,
-            columns=["values"],
+            columns=["values", "nested"],
             group_key="episode",
             order_key="step",
-            column_transforms={"values": mutate},
+            frame_offsets={name: [0, 1, 1, 3, 3] for name in padding},
+            boundary="pad",
+            pad_values=padding,
+            column_transforms={"values": mutate, "nested": mutate_nested},
         )
 
         batched = dataset.__getitems__([0, 1, 0])
         singles = [dataset[index] for index in (0, 1, 0)]
-
-        self.assertEqual(
-            [sample["values"] for sample in singles],
-            [sample["values"] for sample in batched],
-        )
+        expected = [
+            [[0, 99], [1, 99], [1, 99], [-1, 99], [-1, 99]],
+            [[1, 99], [2, 99], [2, 99], [-1, 99], [-1, 99]],
+            [[0, 99], [1, 99], [1, 99], [-1, 99], [-1, 99]],
+        ]
+        for samples in (batched, singles):
+            self.assertEqual(expected, [sample["values"] for sample in samples])
+            self.assertEqual(
+                [[[value] for value in window] for window in expected],
+                [sample["nested"] for sample in samples],
+            )
+        self.assertEqual({"values": [-1], "nested": [[-1]]}, padding)
 
     def test_pad_tail_repeats_last_row_and_marks_real_padding(self):
         dataset = self._dataset(