This paper presents a benchmarking study comparing Kolmogorov-Arnold Networks (KANs) and Multi-Layer Perceptrons (MLPs) on real-world tabular datasets. KANs, inspired by the Kolmogorov-Arnold theorem, offer an interpretable framework with learnable activation functions on edges rather than nodes, allowing for more flexible and adaptive modeling. The study evaluates KANs and MLPs on ten datasets from the UCI Machine Learning Repository, including tasks such as spam detection, diabetes prediction, and poker hand classification. Results show that KANs achieve superior or comparable accuracy and F1 scores, particularly in datasets with many instances, indicating their effectiveness in handling complex data. However, KANs require higher computational resources compared to MLPs. The study also highlights that KANs perform more operations than MLPs, even with similar parameter counts, due to their richer activation functions. The experiments assess task performance, training time, and computational efficiency. KANs consistently outperform MLPs in accuracy and F1 scores, especially in large-scale datasets like Poker and Musk. While precision and recall show similar performance, KANs demonstrate better accuracy in most cases. The study concludes that KANs are a promising alternative to MLPs, particularly for complex datasets, though they come with higher computational costs. Future research could explore KANs in regression tasks and diverse data types to enhance their real-world applicability.This paper presents a benchmarking study comparing Kolmogorov-Arnold Networks (KANs) and Multi-Layer Perceptrons (MLPs) on real-world tabular datasets. KANs, inspired by the Kolmogorov-Arnold theorem, offer an interpretable framework with learnable activation functions on edges rather than nodes, allowing for more flexible and adaptive modeling. The study evaluates KANs and MLPs on ten datasets from the UCI Machine Learning Repository, including tasks such as spam detection, diabetes prediction, and poker hand classification. Results show that KANs achieve superior or comparable accuracy and F1 scores, particularly in datasets with many instances, indicating their effectiveness in handling complex data. However, KANs require higher computational resources compared to MLPs. The study also highlights that KANs perform more operations than MLPs, even with similar parameter counts, due to their richer activation functions. The experiments assess task performance, training time, and computational efficiency. KANs consistently outperform MLPs in accuracy and F1 scores, especially in large-scale datasets like Poker and Musk. While precision and recall show similar performance, KANs demonstrate better accuracy in most cases. The study concludes that KANs are a promising alternative to MLPs, particularly for complex datasets, though they come with higher computational costs. Future research could explore KANs in regression tasks and diverse data types to enhance their real-world applicability.