In natural communication, people often reframe the same idea depending on context, audience, or intent. Capturing this fluidity is central to paraphrase detection, which aims to train models that recognize semantic equivalence beyond surface-level wording. Teaching machines to achieve this remains a core challenge in natural language processing and provides motivation for this study. This work applies two transformer encoders, BERT-base-uncased and RoBERTa-large, to the Microsoft Research Paraphrase Corpus. Sentence embeddings are extracted and compared using cosine similarity and Euclidean distance, with paraphrases labeled at a fixed threshold of 0.70. Performance is evaluated by ROC AUC, accuracy, and balanced accuracy. Results show that BERT with Euclidean distance achieves the highest ROC AUC (0.7858) and best-balanced accuracy (0.7018), while RoBERTa with cosine similarity yields the top accuracy (0.7477), with ROC AUC of 0.7782 and balanced accuracy of 0.6939. These findings demonstrate that both model architecture and similarity measure strongly affect paraphrase detection, underscoring the importance of systematic comparison when addressing semantic similarity tasks.