This paper proposes a novel approach to latent factor conditional asset pricing models using Kolmogorov-Arnold Networks (KANs) as the basis for autoencoders. Unlike traditional methods that rely on Multilayer Perceptrons (MLPs) with ReLU activations, the KAN-based autoencoder offers enhanced flexibility in modeling nonlinear relationships between asset characteristics and factor exposures, while also improving interpretability. The authors demonstrate through empirical backtesting that their model outperforms MLP-based counterparts in explaining cross-sectional risk exposures. Additionally, long-short portfolios constructed using the model's predictions achieve higher Sharpe ratios, indicating practical value for investment management. The work highlights the potential of KANs in financial applications, particularly for asset pricing and portfolio construction.