High-entropy alloys (HEAs) require interatomic potentials that remain accurate across diverse local chemical environments while enabling large-scale simulations. We develop and compare two complementary, interpretable machine-learning potentials for a five-element Ni--Co--Ti--Zr--Hf system, trained on density functional theory (DFT) energies and forces from VASP molecular dynamics trajectories and evaluated by strict out-of-sample testing. KAN-EAM: We preserve the embedded-atom method (EAM) decomposition and replace the constitutive functions ρ, φ, and F(φ) with trainable one-dimensional functions implemented using Kolmogorov--Arnold Networks (KAN). This physics-informed parameterization maintains the learned components' direct interpretability and enables smoothness regularization to stabilize derivatives and enhance transferability. Trained on more than 9,000 DFT frames (8Å cutoff), KAN-EAM is tested on three independent VASP trajectories excluded from training (1500 frames total), achieving test energy RMSE = $9.39×10-3 eV/atom, and force RMSE = 0.2068 eV/Å. KAN-SNAP: We develop a nonlinear variant of the Spectral Neighbor Analysis Potential (SNAP) by replacing the standard linear bispectrum model with element-conditioned one-dimensional KAN functions. The per-atom energy is expressed as E=\sum_i\sum_k ft_i k(Bik), where Bik are bispectrum components and ti denotes the chemical species. Forces are computed analytically by applying the chain rule to the bispectrum derivatives provided by LAMMPS (snad/atom), which allows the model to be trained directly on forces while keeping the learned relations from bispectrum components B to per-atom energies Ei physically interpretable. On an independent test set, the model achieves RMSE$(ΔE/atom) = 6.16×10-3 eV/atom and force RMSE = 0.284 eV/Å indicating promising generalization across compositions and loading conditions. Together, these results highlight (i) a physics-based, interpretable route to EAM-quality scaling via KAN-learned 1D constitutive laws and (ii) a nonlinear, descriptor-driven route to improved expressivity in SNAP, motivating future work on force transferability and robust training across diverse HEA environments.