diff --git a/spacy/tests/pipeline/test_tagger.py b/spacy/tests/pipeline/test_tagger.py
index fdc0a17d5..c717794d8 100644
--- a/spacy/tests/pipeline/test_tagger.py
+++ b/spacy/tests/pipeline/test_tagger.py
@@ -1,5 +1,5 @@
 import pytest
-from numpy.testing import assert_equal
+from numpy.testing import assert_equal, assert_array_almost_equal
 from spacy.attrs import TAG
 
 from spacy import util
@@ -68,7 +68,6 @@ PARTIAL_DATA = [
 
 
 def test_label_smoothing():
-    util.fix_random_seed()
     nlp = Language()
     tagger_no_ls = nlp.add_pipe(
         "tagger", "no_label_smoothing", config=dict(label_smoothing=False)
@@ -83,14 +82,14 @@ def test_label_smoothing():
         tagger_ls.add_label(tag)
     for t in TRAIN_DATA:
         train_examples.append(Example.from_dict(nlp.make_doc(t[0]), t[1]))
-    optimizer = nlp.initialize(get_examples=lambda: train_examples)
-    for i in range(5):
-        losses = {}
-        nlp.update(train_examples, sgd=optimizer, losses=losses)
-    assert losses == {
-        "no_label_smoothing": 1.4892945885658264,
-        "label_smoothing": 1.1432453989982605,
-    }
+
+    nlp.initialize(get_examples=lambda: train_examples)
+    tag_scores, bp_tag_scores = tagger_ls.model.begin_update(
+        [eg.predicted for eg in train_examples]
+    )
+    no_ls_probs = tagger_no_ls.get_loss(train_examples, tag_scores)[1][0]
+    ls_probs = tagger_ls.get_loss(train_examples, tag_scores)[1][0]
+    assert_array_almost_equal((ls_probs - no_ls_probs)[0], [0.05, -0.025, -0.025])
 
 
 def test_no_label():