[2/2] How it works: two coupled implicit neural representations parameterize the permittivity contrast and induced current, jointly optimized under Lippmann-Schwinger physics; the score model injects a learned prior gradient via the chain rule, handling strong multiple scattering

[2/2] The theory is comprehensive: convergence analysis covering single-block and multi-block implementations plus annealing schedules — under transparent assumptions, even for non-log-concave posteriors and imperfect learned score models.

[3/3] Reconstructions stay sharp under severe undersampling. We also characterize the bias from likelihood guidance and propose a ground-truth-free metric for posterior sampling quality.