diff --git a/tests/experimental/test_minillm_trainer.py b/tests/experimental/test_minillm_trainer.py
index 59c0be9f306..619e705dd46 100644
--- a/tests/experimental/test_minillm_trainer.py
+++ b/tests/experimental/test_minillm_trainer.py
@@ -12,6 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import types
+
 import pytest
 import torch
 from datasets import DatasetDict, load_dataset
@@ -80,3 +82,64 @@ def test_init_with_eval_dataset(self, eval_dataset_type):
             assert set(trainer.eval_dataset.keys()) == {"data1", "data2"}
         else:
             assert trainer.eval_dataset is eval_dataset
+
+
+class TestMiniLLMComputeAdvantage:
+    def test_length_normalization_ignores_padding(self):
+        """A constant reward of 1.0 with gamma=1.0 must give an advantage of exactly 1.0 at every valid position,
+        however long the batch is padded. Counting padded slots toward the length, as the old 1e-4 fill did, lowers the
+        first position of a short completion (0.9514 for a single valid token padded to 512), and a floor on the
+        discounted length would deflate late positions once gamma < 1."""
+        for gamma in (1.0, 0.9):
+            stub = types.SimpleNamespace(gamma=gamma, length_normalization=True)
+            for length in (1, 128, 512):
+                mask = torch.zeros(1, 512)
+                mask[0, :length] = 1
+                advantages = MiniLLMTrainer._compute_advantage(stub, torch.zeros(1, 512), torch.ones(1, 512), mask)
+                # With gamma < 1 the discounted length of a late position is tiny (0.9**300 is 1.9e-14), so any floor
+                # on the denominator would deflate it; the expected value is exactly the reward at every position.
+                torch.testing.assert_close(advantages[0, :length], torch.ones(length))
+                torch.testing.assert_close(advantages[0, length:], torch.zeros(512 - length))
+
+            advantages = MiniLLMTrainer._compute_advantage(
+                stub, torch.zeros(1, 8), torch.ones(1, 8), torch.zeros(1, 8)
+            )
+            assert torch.isfinite(advantages).all()
+            torch.testing.assert_close(advantages, torch.zeros(1, 8))
+
+
+class TestMiniLLMComputeLossMask:
+    def test_advantage_mask_excludes_padded_completion_tokens(self, monkeypatch):
+        """The mask handed to `_compute_advantage` must come from the -100 fill in `labels`; built from `input_ids`
+        it was all ones, so padded slots kept a reward and leaked into every earlier advantage."""
+
+        class DummyModel:
+            def eval(self):
+                return self
+
+            def __call__(self, input_ids, attention_mask, use_cache):
+                return types.SimpleNamespace(logits=torch.zeros(*input_ids.shape, 10))
+
+        trainer = object.__new__(MiniLLMTrainer)
+        trainer.teacher_model = DummyModel()
+        trainer.kd_temperature = 1.0
+        trainer.rkl_advantage = True
+
+        completion_mask = torch.tensor([[1, 1, 0, 0], [1, 1, 1, 1]])
+        inputs = {
+            "prompt_ids": torch.randint(1, 10, (2, 3)),
+            "prompt_mask": torch.ones(2, 3, dtype=torch.long),
+            "completion_ids": torch.randint(1, 10, (2, 4)),
+            "completion_mask": completion_mask,
+        }
+
+        class AdvantageCaptured(Exception):
+            pass
+
+        def spy(**kwargs):
+            assert torch.equal(kwargs["mask"], completion_mask.bool())
+            raise AdvantageCaptured
+
+        monkeypatch.setattr(trainer, "_compute_advantage", spy)
+        with pytest.raises(AdvantageCaptured):
+            trainer.compute_loss(DummyModel(), inputs)
diff --git a/trl/experimental/minillm/minillm_trainer.py b/trl/experimental/minillm/minillm_trainer.py
index 646f1355426..e37a8af5882 100644
--- a/trl/experimental/minillm/minillm_trainer.py
+++ b/trl/experimental/minillm/minillm_trainer.py
@@ -336,10 +336,12 @@ def _compute_advantage(
             advantages = advantages.flip(1).cumsum(dim=1).flip(1)
 
             if self.length_normalization:
-                mask = torch.where(mask < 0.5, 1e-4, mask)
-                lengths = mask * gamma_pow
-                lengths = lengths.flip(1).cumsum(dim=1).flip(1)
-                advantages = advantages / lengths
+                # Only unmasked positions count toward the discounted length, so the advantage of a completion does
+                # not depend on how much padding the batch carries. A length is zero only where every later position
+                # is masked, and there the numerator is zero too, so dividing by one keeps that tail at zero without
+                # touching valid positions whose discounted length is small.
+                lengths = (mask * gamma_pow).flip(1).cumsum(dim=1).flip(1)
+                advantages = advantages / lengths.masked_fill(lengths == 0, 1.0)
         else:
             advantages = rewards
 
@@ -380,7 +382,9 @@ def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=N
             teacher_log_probs, dim=-1, index=shifted_labels.unsqueeze(-1)
         ).squeeze(-1)
 
-        mask = shifted_labels != -100
+        # `input_ids` never holds -100; the padding fill lives in `labels`, so the mask has to come from there. The
+        # gather above keeps indexing `input_ids`, since -100 is not a valid index.
+        mask = labels[:, prompt_lengths:] != -100
 
         if self.rkl_advantage:
             reverse_kl_advantage = self._compute_advantage(