Paraphrase detection is one of the challenging tasks in NLP that forms the core of semantic search, determination of duplicate contents, and answering questions. This paper fine-tunes and evaluates three such state-of-the-art transformer models-Roberta, MPNet, and BERT on various datasets that includes Microsoft Research Paraphrase Corpus, Quora Question Pairs, PAWS-X Wiki, and Twitter Paraphrase Corpus datasets. It presents a strong evaluation across different linguistic structures, paraphrase patterns, and domain-specific challenges. Roberta constantly showed high performance in this classification task, thus it may be helpful for high precision-oriented tasks, while MPNet demonstrates superiority in inference time and hence is suitable for low-latency tasks. A thorough error analysis with focus on the linguistic complexity types of idiomatic expressions, negation, and semantic ambiguity revealed strengths and weaknesses of the models. Besides, SHAP and LIME explainability techniques were applied to provide model interpretations and help gain more insights into model predictions by increasing transparency and model trustworthiness. This work sets benchmarks and gives insights into further research in paraphrase detection models that can be useful in relevant applications tuned to particular requirements.