diff --git a/config.yaml b/config.yaml
index 7db0c1f..9104327 100644
--- a/config.yaml
+++ b/config.yaml
@@ -30,4 +30,6 @@ training:
 
 hydra:
   run:
-    dir: logs
\ No newline at end of file
+    dir: logs
+
+run_additional_inference: True  # Set to false to not run multipass inference.
\ No newline at end of file
diff --git a/main.py b/main.py
index c2ddc9e..ba0f1ca 100644
--- a/main.py
+++ b/main.py
@@ -57,6 +57,12 @@ def validate_or_test(opt, model, partition, epoch=None):
             scalar_outputs = model.forward_downstream_classification_model(
                 inputs, labels
             )
+
+            if opt.run_additional_inference:
+                scalar_outputs = model.forward_downstream_multi_pass(
+                    inputs, labels, scalar_outputs=scalar_outputs
+                )
+                
             test_results = utils.log_results(
                 test_results, scalar_outputs, num_steps_per_epoch
             )
diff --git a/src/ff_mnist.py b/src/ff_mnist.py
index d741cd8..d987c4f 100644
--- a/src/ff_mnist.py
+++ b/src/ff_mnist.py
@@ -12,7 +12,7 @@ def __init__(self, opt, partition, num_classes=10):
         self.uniform_label = torch.ones(self.num_classes) / self.num_classes
 
     def __getitem__(self, index):
-        pos_sample, neg_sample, neutral_sample, class_label = self._generate_sample(
+        pos_sample, neg_sample, neutral_sample, all_sample, class_label = self._generate_sample(
             index
         )
 
@@ -20,6 +20,7 @@ def __getitem__(self, index):
             "pos_images": pos_sample,
             "neg_images": neg_sample,
             "neutral_sample": neutral_sample,
+            "all_sample": all_sample
         }
         labels = {"class_labels": class_label}
         return inputs, labels
@@ -32,7 +33,7 @@ def _get_pos_sample(self, sample, class_label):
             torch.tensor(class_label), num_classes=self.num_classes
         )
         pos_sample = sample.clone()
-        pos_sample[:, 0, : self.num_classes] = one_hot_label
+        pos_sample[0, 0, : self.num_classes] = one_hot_label
         return pos_sample
 
     def _get_neg_sample(self, sample, class_label):
@@ -44,12 +45,21 @@ def _get_neg_sample(self, sample, class_label):
             torch.tensor(wrong_class_label), num_classes=self.num_classes
         )
         neg_sample = sample.clone()
-        neg_sample[:, 0, : self.num_classes] = one_hot_label
+        neg_sample[0, 0, : self.num_classes] = one_hot_label
         return neg_sample
 
     def _get_neutral_sample(self, z):
-        z[:, 0, : self.num_classes] = self.uniform_label
+        z[0, 0, : self.num_classes] = self.uniform_label
         return z
+    
+    def _get_all_sample(self, sample):
+        all_samples = torch.zeros((self.num_classes, sample.shape[0], sample.shape[1], sample.shape[2]))
+        for i in range(self.num_classes):
+            all_samples[i, :, :, :] = sample.clone()
+            one_hot_label = torch.nn.functional.one_hot(
+            torch.tensor(i), num_classes=self.num_classes)
+            all_samples[i, 0, 0, : self.num_classes] = one_hot_label.clone()
+        return all_samples
 
     def _generate_sample(self, index):
         # Get MNIST sample.
@@ -57,4 +67,5 @@ def _generate_sample(self, index):
         pos_sample = self._get_pos_sample(sample, class_label)
         neg_sample = self._get_neg_sample(sample, class_label)
         neutral_sample = self._get_neutral_sample(sample)
-        return pos_sample, neg_sample, neutral_sample, class_label
+        all_sample = self._get_all_sample(sample)
+        return pos_sample, neg_sample, neutral_sample, all_sample, class_label
diff --git a/src/ff_model.py b/src/ff_model.py
index 732e639..2895fc6 100644
--- a/src/ff_model.py
+++ b/src/ff_model.py
@@ -118,6 +118,50 @@ def forward(self, inputs, labels):
             inputs, labels, scalar_outputs=scalar_outputs
         )
 
+        if self.opt.run_additional_inference:
+            scalar_outputs = self.forward_downstream_multi_pass(
+                inputs, labels, scalar_outputs=scalar_outputs
+            )
+
+        return scalar_outputs
+    
+    def forward_downstream_multi_pass(
+        self, inputs, labels, scalar_outputs=None,
+    ):
+        if scalar_outputs is None:
+            scalar_outputs = {
+                "Loss": torch.zeros(1, device=self.opt.device),
+            }
+
+        z_all = inputs["all_sample"]
+        z_all = z_all.reshape(z_all.shape[0], z_all.shape[1], -1)
+        ssq_all = []
+        for class_num in range(z_all.shape[1]):
+            z = z_all[:, class_num, :]
+            z = self._layer_norm(z)
+            input_classification_model = []
+
+            with torch.no_grad():
+                for idx, layer in enumerate(self.model):
+                    z = layer(z)
+                    z = self.act_fn.apply(z)
+                    z_unnorm = z.clone()
+                    z = self._layer_norm(z)
+
+                    if idx >= 1:
+                        # print(z.shape)
+                        input_classification_model.append(z_unnorm)
+
+            input_classification_model = torch.concat(input_classification_model, dim=-1)
+            ssq = torch.sum(input_classification_model ** 2, dim=-1)
+            ssq_all.append(ssq)
+        ssq_all = torch.stack(ssq_all, dim=-1)
+        
+        classification_accuracy = utils.get_accuracy(
+            self.opt, ssq_all.data, labels["class_labels"]
+        )
+
+        scalar_outputs["multi_pass_classification_accuracy"] = classification_accuracy
         return scalar_outputs
 
     def forward_downstream_classification_model(