Abstract: Non-convex regularization has great potential in highly squint synthetic aperture radar (SAR) sparse imaging due to its theoretical advantage in super- resolution and amplitude fidelity. Aiming at the problem that regularization parameters are extremely sensitive to model error and scene features, especially in highly squint scenarios where severe range-azimuth coupling and low signal-to-noise ratio (SNR) make the selection of optimal parameters exceptionally difficult. A gradient-guided adaptive non-convex regularization framework was proposed. Utilizing the principle of bi-level optimization, the subgradient of the reconstruction residual was computed with respect to the generalized minimax concave (GMC) norm. The algorithm could automatically search and adjust the optimal hyper parameters during iterative reconstruction. Simulation and experimental results show that, compared to sparse reconstruction algorithms with fixed hyperparameters, the proposed method improved peak signal-to-noise ratio (PSNR) by more than 4% and reduced normalized mean squared error (NMSE) by more than 35%. This solves the algorithm's dependence on manual parameter tuning and improves reconstruction accuracy in highly squint scenarios.