diff --git a/areal/utils/data.py b/areal/utils/data.py
index 1560742a1a..8d16752d17 100644
--- a/areal/utils/data.py
+++ b/areal/utils/data.py
@@ -1423,6 +1423,11 @@ def _build_group_slices(
                 slices.append(slice(offset, offset + sz))
                 offset += sz
             return slices
+        if bs % self.group_size != 0:
+            raise ValueError(
+                f"batch size ({bs}) must be divisible by group_size "
+                f"({self.group_size}) when group_sizes is not provided"
+            )
         return [
             slice(i * self.group_size, (i + 1) * self.group_size)
             for i in range(bs // self.group_size)
diff --git a/tests/test_reward_norm_variable_group.py b/tests/test_reward_norm_variable_group.py
index c6ce65232f..8b89bff5dc 100644
--- a/tests/test_reward_norm_variable_group.py
+++ b/tests/test_reward_norm_variable_group.py
@@ -143,6 +143,15 @@ def test_group_sizes_non_positive_raises():
         norm(x, group_sizes=[4, 0])
 
 
+def test_group_sizes_none_rejects_non_divisible_group_batch():
+    """Implicit fixed-size grouping must not leave a partial tail unnormalized."""
+    norm = Normalization(_group_norm_config())  # group_size=2
+    x = torch.tensor([0.0, 1.0, 2.0], dtype=torch.float32)  # bs=3
+
+    with pytest.raises(ValueError, match="divisible by group_size"):
+        norm(x)
+
+
 def test_adv_norm_style_2d_variable_groups_normalize_per_group():
     """Token-level (2D) advantages normalize per variable-size group on dim 0.