diff --git a/atari/mingpt/model_atari.py b/atari/mingpt/model_atari.py
index 8428811c..68219e2d 100644
--- a/atari/mingpt/model_atari.py
+++ b/atari/mingpt/model_atari.py
@@ -265,7 +265,7 @@ def forward(self, states, actions, targets=None, rtgs=None, timesteps=None):
         if actions is not None and self.model_type == 'reward_conditioned':
             logits = logits[:, 1::3, :] # only keep predictions from state_embeddings
         elif actions is None and self.model_type == 'reward_conditioned':
-            logits = logits[:, 1:, :]
+            logits = logits[:, 1::2, :]
         elif actions is not None and self.model_type == 'naive':
             logits = logits[:, ::2, :] # only keep predictions from state_embeddings
         elif actions is None and self.model_type == 'naive':