Satellite aerosol data assimilation has traditionally relied on retrieved aerosol products, introducing a "retrieve-then-assimilate" pathway that can propagate retrieval uncertainties and impose assumptions inconsistent with atmospheric models. Direct assimilation of satellite shortwave observations provides a more physically consistent alternative by constraining model states directly in observation space, but its application remains challenging because nonlinear radiative transfer calculations and high-dimensional state optimization are computationally demanding. Here we develop LARDAS (Latent-space Aerosol–Radiance Data Assimilation System), a differentiable latent-space framework that directly assimilates multi-band satellite shortwave reflectance to jointly constrain aerosol optical depth (AOD) and land-surface white-sky albedo. LARDAS combines variational autoencoder-based latent representations of aerosol and surface fields with a differentiable neural radiative transfer emulator, enabling gradient-based optimization from observation space to atmospheric state space. Five-band MODIS top-of-atmosphere reflectances are assimilated through an integrated observation operator consisting of latent decoders, an aerosol refinement network, and band-specific neural radiative transfer surrogates trained against VLIDORT. The proposed framework reduces the control space from approximately 2×5×141×141 physical variables to 256 latent variables while maintaining high reconstruction fidelity (median fraction skill scores >0.95 for aerosol fields and >0.98 for surface albedo). The neural radiative transfer emulator reproduces VLIDORT reflectance simulations with correlations exceeding 0.999 and RMSE below 0.003. Controlled experiments demonstrate that LARDAS can recover aerosol and surface states from multi-band reflectance observations. Real-data assimilation experiments using MODIS observations over eastern China show that LARDAS improves independent AERONET AOD evaluation relative to the model background and aerosol-product assimilation, while achieving comparable skill to physical-space radiance assimilation at approximately three orders of magnitude lower online computational cost under the operationally feasible CPU/GPU deployment configurations used here. Independent CERES validation further demonstrates that joint aerosol–surface optimization improves simulated shortwave radiation, particularly surface upward shortwave flux, where reflectance assimilation substantially reduces biases compared with aerosol-only assimilation. These results demonstrate that latent-space observation-to-state inference provides an efficient and physically consistent pathway toward next-generation satellite aerosol assimilation systems.