Understanding the intricate relationships between cis-regulatory elements and gene expression is crucial for decoding genetic regulation in ecological systems. In this study, we introduce a novel application of Kolmogorov-Arnold Networks (KANs) for predicting gene expression across diverse plant species, including Arabidopsis thaliana, Solanum lycopersicum, Sorghum bicolor, and Zea mays. Our model, named KAN-Mixer, utilizes k-mers as inductive biases to capture biologically relevant patterns in nucleotide sequences. By employing token embeddings and mixer architectures, KAN-Mixer enhances both the interpretability and usability of KANs. Our results indicate that KAN-Mixer achieves comparable accuracy to ConvNet-based approaches while offering superior interpretability, making it a robust tool for ecological data analysis in a variety of environments.