diff --git a/submission_runner.py b/submission_runner.py
index d92732145..7f6150d19 100644
--- a/submission_runner.py
+++ b/submission_runner.py
@@ -580,6 +580,10 @@ def score_submission_on_workload(workload: spec.Workload,
     rng = prng.PRNGKey(rng_seed)
     # If the submission is responsible for tuning itself, we only need to run it
     # once and return the total time.
+    if log_dir is not None:
+      log_dir = os.path.join(log_dir, 'trial_1')
+      logging.info(f'Creating directory at {log_dir}.')
+      logger_utils.makedir(log_dir)
     with profiler.profile('Train'):
       score, _ = train_once(
           workload, global_batch_size, global_eval_batch_size,