Interference cancellation (IC) is crucial for sustaining high-performance massive MIMO downlink receivers at cell edges, where capacity is severely limited by strong co-channel interference. Achieving Bayes-optimal recovery of this interference is fundamentally challenging, as it typically requires explicit prior knowledge of the interfering modulation formats. To overcome this limitation, we propose a novel, modulation-agnostic interference estimation approach that unfolds the Approximate Message Passing (AMP) framework, replacing its standard static denoiser with a dynamic Empirical Bayes (EB) component. In this synergistic architecture, AMP systematically decouples the spatial MIMO channel to provide the scalar environment required by EB, while EB adaptively learns the prior distribution and infers the signal without any pre-existing modulation knowledge. This approach completely circumvents the need for brute-force exhaustive searches, rendering it significantly more efficient than conventional model selection methods, while exhibiting superior computational scalability compared to Expectation-Maximization (EM) based AMP. We integrate this EB-AMP estimator into a comprehensive joint detection framework that iteratively alternates among data detection, signal subtraction, interference recovery, and interference cancellation. Numerical simulations demonstrate that under the assumptions of perfect channel state information (CSI), i.i.d. Rayleigh fading, and a single interfering cell, the proposed IC receiver effectively suppresses unknown interference at power levels comparable to the desired data signal, achieving an estimation accuracy that closely approaches the theoretical interference-free bound.