diff --git a/torchseq/metric_hooks/sep_ae.py b/torchseq/metric_hooks/sep_ae.py
index 4ea6271..14ccd54 100644
--- a/torchseq/metric_hooks/sep_ae.py
+++ b/torchseq/metric_hooks/sep_ae.py
@@ -855,6 +855,19 @@ def eval_gen_codepred_diversity(
 
             topk_outputs.append(output)
 
+        # calculate p-BLEU (Cao and Wan, 2020)
+        # p-BLEU = sum_i, sum_{j neq i} BLEU(yi, yj) / k * (k-1)
+        pbleu_scores = []
+        for i in range(top_k):
+            for j in range(top_k):
+                if i == j:
+                    continue
+                this_bleu = sacrebleu.corpus_bleu(
+                    topk_outputs[i], list(zip(*[[x] for x in topk_outputs[j]])), lowercase=True
+                ).score
+                pbleu_scores.append(this_bleu)
+        scores["pbleu"] = np.mean(pbleu_scores)
+
         agent.config.eval.data["sample_outputs"] = sample_outputs
 
         return scores, topk_outputs